{
 "schema": 1,
 "generated": "2026-10-28T08:15:34Z",
 "kit_version": "0.2.0",
 "confidence_level": 0.95,
 "n_runs": 74,
 "rows": [
  {"model": "convnext-v2", "task": "ptbxl-multilabel", "averaging": "macro-label", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.6836555125876786, "std": null, "lo": null, "hi": null, "values": [0.6836555125876786]},
    "auroc": {"n": 1, "mean": 0.9046229653923364, "std": null, "lo": null, "hi": null, "values": [0.9046229653923364]},
    "precision": {"n": 1, "mean": 0.5983326116189117, "std": null, "lo": null, "hi": null, "values": [0.5983326116189117]},
    "recall": {"n": 1, "mean": 0.8125111970853907, "std": null, "lo": null, "hi": null, "values": [0.8125111970853907]},
    "exact_match_accuracy": {"n": 1, "mean": 0.4749768303985171, "std": null, "lo": null, "hi": null, "values": [0.4749768303985171]},
    "label_accuracy": {"n": 1, "mean": 0.8319740500463392, "std": null, "lo": null, "hi": null, "values": [0.8319740500463392]},
    "ece": {"n": 1, "mean": 0.1136533338720338, "std": null, "lo": null, "hi": null, "values": [0.1136533338720338]},
    "adaptive_ece": {"n": 1, "mean": 0.112838192667303, "std": null, "lo": null, "hi": null, "values": [0.112838192667303]},
    "mce": {"n": 1, "mean": 0.5090338505547622, "std": null, "lo": null, "hi": null, "values": [0.5090338505547622]},
    "brier": {"n": 1, "mean": 0.1186033047582481, "std": null, "lo": null, "hi": null, "values": [0.1186033047582481]},
    "temperature": {"n": 1, "mean": 1.1924841823167027, "std": null, "lo": null, "hi": null, "values": [1.1924841823167027]},
    "ece_temp": {"n": 1, "mean": 0.1194438551539663, "std": null, "lo": null, "hi": null, "values": [0.1194438551539663]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.1178279409145372, "std": null, "lo": null, "hi": null, "values": [0.1178279409145372]},
    "mce_temp": {"n": 1, "mean": 0.4993053984128537, "std": null, "lo": null, "hi": null, "values": [0.4993053984128537]},
    "brier_temp": {"n": 1, "mean": 0.1176560087907504, "std": null, "lo": null, "hi": null, "values": [0.1176560087907504]},
    "ece_platt": {"n": 1, "mean": 0.0219804136674808, "std": null, "lo": null, "hi": null, "values": [0.0219804136674808]},
    "adaptive_ece_platt": {"n": 1, "mean": 0.0187620787692125, "std": null, "lo": null, "hi": null, "values": [0.0187620787692125]},
    "mce_platt": {"n": 1, "mean": 0.215992048158349, "std": null, "lo": null, "hi": null, "values": [0.215992048158349]},
    "brier_platt": {"n": 1, "mean": 0.0903370039718089, "std": null, "lo": null, "hi": null, "values": [0.0903370039718089]}
   }},
  {"model": "crnn", "task": "ptbxl-multilabel", "averaging": "macro-label", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.6783442211940882, "std": null, "lo": null, "hi": null, "values": [0.6783442211940882]},
    "auroc": {"n": 1, "mean": 0.8887733037895561, "std": null, "lo": null, "hi": null, "values": [0.8887733037895561]},
    "precision": {"n": 1, "mean": 0.6133965046005653, "std": null, "lo": null, "hi": null, "values": [0.6133965046005653]},
    "recall": {"n": 1, "mean": 0.7982512080315113, "std": null, "lo": null, "hi": null, "values": [0.7982512080315113]},
    "exact_match_accuracy": {"n": 1, "mean": 0.4601482854494902, "std": null, "lo": null, "hi": null, "values": [0.4601482854494902]},
    "label_accuracy": {"n": 1, "mean": 0.8234476367006488, "std": null, "lo": null, "hi": null, "values": [0.8234476367006488]},
    "ece": {"n": 1, "mean": 0.114738893915349, "std": null, "lo": null, "hi": null, "values": [0.114738893915349]},
    "adaptive_ece": {"n": 1, "mean": 0.1166931232796, "std": null, "lo": null, "hi": null, "values": [0.1166931232796]},
    "mce": {"n": 1, "mean": 0.4685251498726052, "std": null, "lo": null, "hi": null, "values": [0.4685251498726052]},
    "brier": {"n": 1, "mean": 0.1219401985661165, "std": null, "lo": null, "hi": null, "values": [0.1219401985661165]},
    "temperature": {"n": 1, "mean": 1.122087021113016, "std": null, "lo": null, "hi": null, "values": [1.122087021113016]},
    "ece_temp": {"n": 1, "mean": 0.1199398639276995, "std": null, "lo": null, "hi": null, "values": [0.1199398639276995]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.1197398971183403, "std": null, "lo": null, "hi": null, "values": [0.1197398971183403]},
    "mce_temp": {"n": 1, "mean": 0.4636658576701462, "std": null, "lo": null, "hi": null, "values": [0.4636658576701462]},
    "brier_temp": {"n": 1, "mean": 0.1218472356389962, "std": null, "lo": null, "hi": null, "values": [0.1218472356389962]},
    "ece_platt": {"n": 1, "mean": 0.022911038781589, "std": null, "lo": null, "hi": null, "values": [0.022911038781589]},
    "adaptive_ece_platt": {"n": 1, "mean": 0.0194268934984377, "std": null, "lo": null, "hi": null, "values": [0.0194268934984377]},
    "mce_platt": {"n": 1, "mean": 0.1840885840197991, "std": null, "lo": null, "hi": null, "values": [0.1840885840197991]},
    "brier_platt": {"n": 1, "mean": 0.0941588982173711, "std": null, "lo": null, "hi": null, "values": [0.0941588982173711]}
   }},
  {"model": "deep-res-cnn", "task": "ptbxl-multilabel", "averaging": "macro-label", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.71141731721003, "std": null, "lo": null, "hi": null, "values": [0.71141731721003]},
    "auroc": {"n": 1, "mean": 0.9138386574406848, "std": null, "lo": null, "hi": null, "values": [0.9138386574406848]},
    "precision": {"n": 1, "mean": 0.6338555695856429, "std": null, "lo": null, "hi": null, "values": [0.6338555695856429]},
    "recall": {"n": 1, "mean": 0.8413699568600073, "std": null, "lo": null, "hi": null, "values": [0.8413699568600073]},
    "exact_match_accuracy": {"n": 1, "mean": 0.484708063021316, "std": null, "lo": null, "hi": null, "values": [0.484708063021316]},
    "label_accuracy": {"n": 1, "mean": 0.8450417052826691, "std": null, "lo": null, "hi": null, "values": [0.8450417052826691]},
    "ece": {"n": 1, "mean": 0.1042807529250808, "std": null, "lo": null, "hi": null, "values": [0.1042807529250808]},
    "adaptive_ece": {"n": 1, "mean": 0.1058319484250406, "std": null, "lo": null, "hi": null, "values": [0.1058319484250406]},
    "mce": {"n": 1, "mean": 0.5814155797163646, "std": null, "lo": null, "hi": null, "values": [0.5814155797163646]},
    "brier": {"n": 1, "mean": 0.1107647184711946, "std": null, "lo": null, "hi": null, "values": [0.1107647184711946]},
    "temperature": {"n": 1, "mean": 1.1037336764939574, "std": null, "lo": null, "hi": null, "values": [1.1037336764939574]},
    "ece_temp": {"n": 1, "mean": 0.1078821116734296, "std": null, "lo": null, "hi": null, "values": [0.1078821116734296]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.1093326017436802, "std": null, "lo": null, "hi": null, "values": [0.1093326017436802]},
    "mce_temp": {"n": 1, "mean": 0.5819821038141632, "std": null, "lo": null, "hi": null, "values": [0.5819821038141632]},
    "brier_temp": {"n": 1, "mean": 0.1104312737831505, "std": null, "lo": null, "hi": null, "values": [0.1104312737831505]},
    "ece_platt": {"n": 1, "mean": 0.0217050142209892, "std": null, "lo": null, "hi": null, "values": [0.0217050142209892]},
    "adaptive_ece_platt": {"n": 1, "mean": 0.0191303096887456, "std": null, "lo": null, "hi": null, "values": [0.0191303096887456]},
    "mce_platt": {"n": 1, "mean": 0.169849760753422, "std": null, "lo": null, "hi": null, "values": [0.169849760753422]},
    "brier_platt": {"n": 1, "mean": 0.0857052964068146, "std": null, "lo": null, "hi": null, "values": [0.0857052964068146]}
   }},
  {"model": "dualnet", "task": "ptbxl-multilabel", "averaging": "macro-label", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.6694187661441475, "std": null, "lo": null, "hi": null, "values": [0.6694187661441475]},
    "auroc": {"n": 1, "mean": 0.8793312091069717, "std": null, "lo": null, "hi": null, "values": [0.8793312091069717]},
    "precision": {"n": 1, "mean": 0.591103290284677, "std": null, "lo": null, "hi": null, "values": [0.591103290284677]},
    "recall": {"n": 1, "mean": 0.7882356252939061, "std": null, "lo": null, "hi": null, "values": [0.7882356252939061]},
    "exact_match_accuracy": {"n": 1, "mean": 0.4582947173308619, "std": null, "lo": null, "hi": null, "values": [0.4582947173308619]},
    "label_accuracy": {"n": 1, "mean": 0.8240963855421687, "std": null, "lo": null, "hi": null, "values": [0.8240963855421687]},
    "ece": {"n": 1, "mean": 0.1265733918410481, "std": null, "lo": null, "hi": null, "values": [0.1265733918410481]},
    "adaptive_ece": {"n": 1, "mean": 0.1256994881073697, "std": null, "lo": null, "hi": null, "values": [0.1256994881073697]},
    "mce": {"n": 1, "mean": 0.5158997690173942, "std": null, "lo": null, "hi": null, "values": [0.5158997690173942]},
    "brier": {"n": 1, "mean": 0.1275632007136813, "std": null, "lo": null, "hi": null, "values": [0.1275632007136813]},
    "temperature": {"n": 1, "mean": 1.0561231211065574, "std": null, "lo": null, "hi": null, "values": [1.0561231211065574]},
    "ece_temp": {"n": 1, "mean": 0.1285322012411457, "std": null, "lo": null, "hi": null, "values": [0.1285322012411457]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.127510123839397, "std": null, "lo": null, "hi": null, "values": [0.127510123839397]},
    "mce_temp": {"n": 1, "mean": 0.468118289681504, "std": null, "lo": null, "hi": null, "values": [0.468118289681504]},
    "brier_temp": {"n": 1, "mean": 0.1277288369717794, "std": null, "lo": null, "hi": null, "values": [0.1277288369717794]},
    "ece_platt": {"n": 1, "mean": 0.0218936685818278, "std": null, "lo": null, "hi": null, "values": [0.0218936685818278]},
    "adaptive_ece_platt": {"n": 1, "mean": 0.0215177165701792, "std": null, "lo": null, "hi": null, "values": [0.0215177165701792]},
    "mce_platt": {"n": 1, "mean": 0.1916933782298767, "std": null, "lo": null, "hi": null, "values": [0.1916933782298767]},
    "brier_platt": {"n": 1, "mean": 0.0991236393666195, "std": null, "lo": null, "hi": null, "values": [0.0991236393666195]}
   }},
  {"model": "fcn-wang", "task": "ptbxl-multilabel", "averaging": "macro-label", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.6741649144081667, "std": null, "lo": null, "hi": null, "values": [0.6741649144081667]},
    "auroc": {"n": 1, "mean": 0.8922700287384512, "std": null, "lo": null, "hi": null, "values": [0.8922700287384512]},
    "precision": {"n": 1, "mean": 0.5963442343495726, "std": null, "lo": null, "hi": null, "values": [0.5963442343495726]},
    "recall": {"n": 1, "mean": 0.7985799541976532, "std": null, "lo": null, "hi": null, "values": [0.7985799541976532]},
    "exact_match_accuracy": {"n": 1, "mean": 0.448563484708063, "std": null, "lo": null, "hi": null, "values": [0.448563484708063]},
    "label_accuracy": {"n": 1, "mean": 0.8240963855421687, "std": null, "lo": null, "hi": null, "values": [0.8240963855421687]},
    "ece": {"n": 1, "mean": 0.1233192710820924, "std": null, "lo": null, "hi": null, "values": [0.1233192710820924]},
    "adaptive_ece": {"n": 1, "mean": 0.1229633395216154, "std": null, "lo": null, "hi": null, "values": [0.1229633395216154]},
    "mce": {"n": 1, "mean": 0.4777666895973438, "std": null, "lo": null, "hi": null, "values": [0.4777666895973438]},
    "brier": {"n": 1, "mean": 0.1240330967677413, "std": null, "lo": null, "hi": null, "values": [0.1240330967677413]},
    "temperature": {"n": 1, "mean": 1.0476478610137183, "std": null, "lo": null, "hi": null, "values": [1.0476478610137183]},
    "ece_temp": {"n": 1, "mean": 0.1258302443362955, "std": null, "lo": null, "hi": null, "values": [0.1258302443362955]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.1251439327925602, "std": null, "lo": null, "hi": null, "values": [0.1251439327925602]},
    "mce_temp": {"n": 1, "mean": 0.4663412015704958, "std": null, "lo": null, "hi": null, "values": [0.4663412015704958]},
    "brier_temp": {"n": 1, "mean": 0.1241488337818507, "std": null, "lo": null, "hi": null, "values": [0.1241488337818507]},
    "ece_platt": {"n": 1, "mean": 0.024607979876968, "std": null, "lo": null, "hi": null, "values": [0.024607979876968]},
    "adaptive_ece_platt": {"n": 1, "mean": 0.020493644109511, "std": null, "lo": null, "hi": null, "values": [0.020493644109511]},
    "mce_platt": {"n": 1, "mean": 0.8675631367537555, "std": null, "lo": null, "hi": null, "values": [0.8675631367537555]},
    "brier_platt": {"n": 1, "mean": 0.096511165008535, "std": null, "lo": null, "hi": null, "values": [0.096511165008535]}
   }},
  {"model": "gru", "task": "ptbxl-multilabel", "averaging": "macro-label", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.6826187773441816, "std": null, "lo": null, "hi": null, "values": [0.6826187773441816]},
    "auroc": {"n": 1, "mean": 0.8978143189531629, "std": null, "lo": null, "hi": null, "values": [0.8978143189531629]},
    "precision": {"n": 1, "mean": 0.6088327199444035, "std": null, "lo": null, "hi": null, "values": [0.6088327199444035]},
    "recall": {"n": 1, "mean": 0.8097115356760293, "std": null, "lo": null, "hi": null, "values": [0.8097115356760293]},
    "exact_match_accuracy": {"n": 1, "mean": 0.4481000926784059, "std": null, "lo": null, "hi": null, "values": [0.4481000926784059]},
    "label_accuracy": {"n": 1, "mean": 0.8270620945319741, "std": null, "lo": null, "hi": null, "values": [0.8270620945319741]},
    "ece": {"n": 1, "mean": 0.1104631705059132, "std": null, "lo": null, "hi": null, "values": [0.1104631705059132]},
    "adaptive_ece": {"n": 1, "mean": 0.1113376070708459, "std": null, "lo": null, "hi": null, "values": [0.1113376070708459]},
    "mce": {"n": 1, "mean": 0.5800267556045628, "std": null, "lo": null, "hi": null, "values": [0.5800267556045628]},
    "brier": {"n": 1, "mean": 0.1231998865824191, "std": null, "lo": null, "hi": null, "values": [0.1231998865824191]},
    "temperature": {"n": 1, "mean": 1.3122636742806726, "std": null, "lo": null, "hi": null, "values": [1.3122636742806726]},
    "ece_temp": {"n": 1, "mean": 0.1226927174552314, "std": null, "lo": null, "hi": null, "values": [0.1226927174552314]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.1227791505981071, "std": null, "lo": null, "hi": null, "values": [0.1227791505981071]},
    "mce_temp": {"n": 1, "mean": 0.5387246744511376, "std": null, "lo": null, "hi": null, "values": [0.5387246744511376]},
    "brier_temp": {"n": 1, "mean": 0.1221920826151047, "std": null, "lo": null, "hi": null, "values": [0.1221920826151047]},
    "ece_platt": {"n": 1, "mean": 0.0199518113225757, "std": null, "lo": null, "hi": null, "values": [0.0199518113225757]},
    "adaptive_ece_platt": {"n": 1, "mean": 0.0197990291922748, "std": null, "lo": null, "hi": null, "values": [0.0197990291922748]},
    "mce_platt": {"n": 1, "mean": 0.2391889060477695, "std": null, "lo": null, "hi": null, "values": [0.2391889060477695]},
    "brier_platt": {"n": 1, "mean": 0.0931152358005165, "std": null, "lo": null, "hi": null, "values": [0.0931152358005165]}
   }},
  {"model": "inception-time", "task": "ptbxl-multilabel", "averaging": "macro-label", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.699480872831462, "std": null, "lo": null, "hi": null, "values": [0.699480872831462]},
    "auroc": {"n": 1, "mean": 0.9094460086996404, "std": null, "lo": null, "hi": null, "values": [0.9094460086996404]},
    "precision": {"n": 1, "mean": 0.6226984421113425, "std": null, "lo": null, "hi": null, "values": [0.6226984421113425]},
    "recall": {"n": 1, "mean": 0.8252495200378563, "std": null, "lo": null, "hi": null, "values": [0.8252495200378563]},
    "exact_match_accuracy": {"n": 1, "mean": 0.4768303985171455, "std": null, "lo": null, "hi": null, "values": [0.4768303985171455]},
    "label_accuracy": {"n": 1, "mean": 0.8388322520852641, "std": null, "lo": null, "hi": null, "values": [0.8388322520852641]},
    "ece": {"n": 1, "mean": 0.1059434400687409, "std": null, "lo": null, "hi": null, "values": [0.1059434400687409]},
    "adaptive_ece": {"n": 1, "mean": 0.1054644683933401, "std": null, "lo": null, "hi": null, "values": [0.1054644683933401]},
    "mce": {"n": 1, "mean": 0.4949815142644594, "std": null, "lo": null, "hi": null, "values": [0.4949815142644594]},
    "brier": {"n": 1, "mean": 0.1134968244753556, "std": null, "lo": null, "hi": null, "values": [0.1134968244753556]},
    "temperature": {"n": 1, "mean": 1.1336318009496948, "std": null, "lo": null, "hi": null, "values": [1.1336318009496948]},
    "ece_temp": {"n": 1, "mean": 0.1109417566697557, "std": null, "lo": null, "hi": null, "values": [0.1109417566697557]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.1107629771811568, "std": null, "lo": null, "hi": null, "values": [0.1107629771811568]},
    "mce_temp": {"n": 1, "mean": 0.4570772827335472, "std": null, "lo": null, "hi": null, "values": [0.4570772827335472]},
    "brier_temp": {"n": 1, "mean": 0.1133428531598073, "std": null, "lo": null, "hi": null, "values": [0.1133428531598073]},
    "ece_platt": {"n": 1, "mean": 0.0189428286635936, "std": null, "lo": null, "hi": null, "values": [0.0189428286635936]},
    "adaptive_ece_platt": {"n": 1, "mean": 0.0198796440768674, "std": null, "lo": null, "hi": null, "values": [0.0198796440768674]},
    "mce_platt": {"n": 1, "mean": 0.2346383631821044, "std": null, "lo": null, "hi": null, "values": [0.2346383631821044]},
    "brier_platt": {"n": 1, "mean": 0.0889444554575858, "std": null, "lo": null, "hi": null, "values": [0.0889444554575858]}
   }},
  {"model": "kanres", "task": "ptbxl-multilabel", "averaging": "macro-label", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.660078112426514, "std": null, "lo": null, "hi": null, "values": [0.660078112426514]},
    "auroc": {"n": 1, "mean": 0.8810608454694432, "std": null, "lo": null, "hi": null, "values": [0.8810608454694432]},
    "precision": {"n": 1, "mean": 0.582845960201135, "std": null, "lo": null, "hi": null, "values": [0.582845960201135]},
    "recall": {"n": 1, "mean": 0.7878728194781099, "std": null, "lo": null, "hi": null, "values": [0.7878728194781099]},
    "exact_match_accuracy": {"n": 1, "mean": 0.4397590361445783, "std": null, "lo": null, "hi": null, "values": [0.4397590361445783]},
    "label_accuracy": {"n": 1, "mean": 0.8137164040778498, "std": null, "lo": null, "hi": null, "values": [0.8137164040778498]},
    "ece": {"n": 1, "mean": 0.1281792472856868, "std": null, "lo": null, "hi": null, "values": [0.1281792472856868]},
    "adaptive_ece": {"n": 1, "mean": 0.1281360283536963, "std": null, "lo": null, "hi": null, "values": [0.1281360283536963]},
    "mce": {"n": 1, "mean": 0.4923862932726394, "std": null, "lo": null, "hi": null, "values": [0.4923862932726394]},
    "brier": {"n": 1, "mean": 0.1307800752160057, "std": null, "lo": null, "hi": null, "values": [0.1307800752160057]},
    "temperature": {"n": 1, "mean": 1.0384475479154118, "std": null, "lo": null, "hi": null, "values": [1.0384475479154118]},
    "ece_temp": {"n": 1, "mean": 0.13010320877836, "std": null, "lo": null, "hi": null, "values": [0.13010320877836]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.1299757020251169, "std": null, "lo": null, "hi": null, "values": [0.1299757020251169]},
    "mce_temp": {"n": 1, "mean": 0.4749683241033275, "std": null, "lo": null, "hi": null, "values": [0.4749683241033275]},
    "brier_temp": {"n": 1, "mean": 0.1308511022915422, "std": null, "lo": null, "hi": null, "values": [0.1308511022915422]},
    "ece_platt": {"n": 1, "mean": 0.0182417411573242, "std": null, "lo": null, "hi": null, "values": [0.0182417411573242]},
    "adaptive_ece_platt": {"n": 1, "mean": 0.0196254583402976, "std": null, "lo": null, "hi": null, "values": [0.0196254583402976]},
    "mce_platt": {"n": 1, "mean": 0.8276004805876412, "std": null, "lo": null, "hi": null, "values": [0.8276004805876412]},
    "brier_platt": {"n": 1, "mean": 0.1010858851138872, "std": null, "lo": null, "hi": null, "values": [0.1010858851138872]}
   }},
  {"model": "kanres-deep", "task": "ptbxl-multilabel", "averaging": "macro-label", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.6921775120865543, "std": null, "lo": null, "hi": null, "values": [0.6921775120865543]},
    "auroc": {"n": 1, "mean": 0.908224615432137, "std": null, "lo": null, "hi": null, "values": [0.908224615432137]},
    "precision": {"n": 1, "mean": 0.6064870524064403, "std": null, "lo": null, "hi": null, "values": [0.6064870524064403]},
    "recall": {"n": 1, "mean": 0.8389212832680885, "std": null, "lo": null, "hi": null, "values": [0.8389212832680885]},
    "exact_match_accuracy": {"n": 1, "mean": 0.4508804448563485, "std": null, "lo": null, "hi": null, "values": [0.4508804448563485]},
    "label_accuracy": {"n": 1, "mean": 0.8296570898980538, "std": null, "lo": null, "hi": null, "values": [0.8296570898980538]},
    "ece": {"n": 1, "mean": 0.1162611841330434, "std": null, "lo": null, "hi": null, "values": [0.1162611841330434]},
    "adaptive_ece": {"n": 1, "mean": 0.1176758871325727, "std": null, "lo": null, "hi": null, "values": [0.1176758871325727]},
    "mce": {"n": 1, "mean": 0.5546403594315051, "std": null, "lo": null, "hi": null, "values": [0.5546403594315051]},
    "brier": {"n": 1, "mean": 0.1204533555680013, "std": null, "lo": null, "hi": null, "values": [0.1204533555680013]},
    "temperature": {"n": 1, "mean": 1.202602148492475, "std": null, "lo": null, "hi": null, "values": [1.202602148492475]},
    "ece_temp": {"n": 1, "mean": 0.1219431579481337, "std": null, "lo": null, "hi": null, "values": [0.1219431579481337]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.1231818563927447, "std": null, "lo": null, "hi": null, "values": [0.1231818563927447]},
    "mce_temp": {"n": 1, "mean": 0.5074008346612007, "std": null, "lo": null, "hi": null, "values": [0.5074008346612007]},
    "brier_temp": {"n": 1, "mean": 0.1195995165817686, "std": null, "lo": null, "hi": null, "values": [0.1195995165817686]},
    "ece_platt": {"n": 1, "mean": 0.01942882115827, "std": null, "lo": null, "hi": null, "values": [0.01942882115827]},
    "adaptive_ece_platt": {"n": 1, "mean": 0.0187254049695343, "std": null, "lo": null, "hi": null, "values": [0.0187254049695343]},
    "mce_platt": {"n": 1, "mean": 0.208629826252573, "std": null, "lo": null, "hi": null, "values": [0.208629826252573]},
    "brier_platt": {"n": 1, "mean": 0.0889057660599297, "std": null, "lo": null, "hi": null, "values": [0.0889057660599297]}
   }},
  {"model": "lstm", "task": "ptbxl-multilabel", "averaging": "macro-label", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.635546715432133, "std": null, "lo": null, "hi": null, "values": [0.635546715432133]},
    "auroc": {"n": 1, "mean": 0.8588864578639466, "std": null, "lo": null, "hi": null, "values": [0.8588864578639466]},
    "precision": {"n": 1, "mean": 0.5587246602660032, "std": null, "lo": null, "hi": null, "values": [0.5587246602660032]},
    "recall": {"n": 1, "mean": 0.7783460726495168, "std": null, "lo": null, "hi": null, "values": [0.7783460726495168]},
    "exact_match_accuracy": {"n": 1, "mean": 0.4263206672845227, "std": null, "lo": null, "hi": null, "values": [0.4263206672845227]},
    "label_accuracy": {"n": 1, "mean": 0.7939759036144578, "std": null, "lo": null, "hi": null, "values": [0.7939759036144578]},
    "ece": {"n": 1, "mean": 0.1433619246700931, "std": null, "lo": null, "hi": null, "values": [0.1433619246700931]},
    "adaptive_ece": {"n": 1, "mean": 0.1429573426332259, "std": null, "lo": null, "hi": null, "values": [0.1429573426332259]},
    "mce": {"n": 1, "mean": 0.5379172467759678, "std": null, "lo": null, "hi": null, "values": [0.5379172467759678]},
    "brier": {"n": 1, "mean": 0.1459408537836097, "std": null, "lo": null, "hi": null, "values": [0.1459408537836097]},
    "temperature": {"n": 1, "mean": 1.1413708692629787, "std": null, "lo": null, "hi": null, "values": [1.1413708692629787]},
    "ece_temp": {"n": 1, "mean": 0.1464920604603826, "std": null, "lo": null, "hi": null, "values": [0.1464920604603826]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.1467218083215259, "std": null, "lo": null, "hi": null, "values": [0.1467218083215259]},
    "mce_temp": {"n": 1, "mean": 0.5014455357271348, "std": null, "lo": null, "hi": null, "values": [0.5014455357271348]},
    "brier_temp": {"n": 1, "mean": 0.1457052520291737, "std": null, "lo": null, "hi": null, "values": [0.1457052520291737]},
    "ece_platt": {"n": 1, "mean": 0.0197151127418243, "std": null, "lo": null, "hi": null, "values": [0.0197151127418243]},
    "adaptive_ece_platt": {"n": 1, "mean": 0.0192961431251803, "std": null, "lo": null, "hi": null, "values": [0.0192961431251803]},
    "mce_platt": {"n": 1, "mean": 0.3383396894988955, "std": null, "lo": null, "hi": null, "values": [0.3383396894988955]},
    "brier_platt": {"n": 1, "mean": 0.109533919659522, "std": null, "lo": null, "hi": null, "values": [0.109533919659522]}
   }},
  {"model": "medformer", "task": "ptbxl-multilabel", "averaging": "macro-label", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.6441930160714526, "std": null, "lo": null, "hi": null, "values": [0.6441930160714526]},
    "auroc": {"n": 1, "mean": 0.87278831661933, "std": null, "lo": null, "hi": null, "values": [0.87278831661933]},
    "precision": {"n": 1, "mean": 0.5632980625681802, "std": null, "lo": null, "hi": null, "values": [0.5632980625681802]},
    "recall": {"n": 1, "mean": 0.7939463016527516, "std": null, "lo": null, "hi": null, "values": [0.7939463016527516]},
    "exact_match_accuracy": {"n": 1, "mean": 0.4003707136237256, "std": null, "lo": null, "hi": null, "values": [0.4003707136237256]},
    "label_accuracy": {"n": 1, "mean": 0.7951807228915663, "std": null, "lo": null, "hi": null, "values": [0.7951807228915663]},
    "ece": {"n": 1, "mean": 0.141989083396392, "std": null, "lo": null, "hi": null, "values": [0.141989083396392]},
    "adaptive_ece": {"n": 1, "mean": 0.1414012764467335, "std": null, "lo": null, "hi": null, "values": [0.1414012764467335]},
    "mce": {"n": 1, "mean": 0.5640032032949734, "std": null, "lo": null, "hi": null, "values": [0.5640032032949734]},
    "brier": {"n": 1, "mean": 0.141987047239145, "std": null, "lo": null, "hi": null, "values": [0.141987047239145]},
    "temperature": {"n": 1, "mean": 1.1290027973295225, "std": null, "lo": null, "hi": null, "values": [1.1290027973295225]},
    "ece_temp": {"n": 1, "mean": 0.1447412994597556, "std": null, "lo": null, "hi": null, "values": [0.1447412994597556]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.1448554457017289, "std": null, "lo": null, "hi": null, "values": [0.1448554457017289]},
    "mce_temp": {"n": 1, "mean": 0.5417282165069466, "std": null, "lo": null, "hi": null, "values": [0.5417282165069466]},
    "brier_temp": {"n": 1, "mean": 0.1413232825474321, "std": null, "lo": null, "hi": null, "values": [0.1413232825474321]},
    "ece_platt": {"n": 1, "mean": 0.0195256822920072, "std": null, "lo": null, "hi": null, "values": [0.0195256822920072]},
    "adaptive_ece_platt": {"n": 1, "mean": 0.021814845439719, "std": null, "lo": null, "hi": null, "values": [0.021814845439719]},
    "mce_platt": {"n": 1, "mean": 0.1356380890584728, "std": null, "lo": null, "hi": null, "values": [0.1356380890584728]},
    "brier_platt": {"n": 1, "mean": 0.1046209312019128, "std": null, "lo": null, "hi": null, "values": [0.1046209312019128]}
   }},
  {"model": "mlp-wang", "task": "ptbxl-multilabel", "averaging": "macro-label", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.4609866907585623, "std": null, "lo": null, "hi": null, "values": [0.4609866907585623]},
    "auroc": {"n": 1, "mean": 0.6871832469153107, "std": null, "lo": null, "hi": null, "values": [0.6871832469153107]},
    "precision": {"n": 1, "mean": 0.376374067250875, "std": null, "lo": null, "hi": null, "values": [0.376374067250875]},
    "recall": {"n": 1, "mean": 0.6412574729734475, "std": null, "lo": null, "hi": null, "values": [0.6412574729734475]},
    "exact_match_accuracy": {"n": 1, "mean": 0.2919369786839666, "std": null, "lo": null, "hi": null, "values": [0.2919369786839666]},
    "label_accuracy": {"n": 1, "mean": 0.6417052826691381, "std": null, "lo": null, "hi": null, "values": [0.6417052826691381]},
    "ece": {"n": 1, "mean": 0.194836051358111, "std": null, "lo": null, "hi": null, "values": [0.194836051358111]},
    "adaptive_ece": {"n": 1, "mean": 0.1939273188769023, "std": null, "lo": null, "hi": null, "values": [0.1939273188769023]},
    "mce": {"n": 1, "mean": 0.9878316521644592, "std": null, "lo": null, "hi": null, "values": [0.9878316521644592]},
    "brier": {"n": 1, "mean": 0.2129425172861202, "std": null, "lo": null, "hi": null, "values": [0.2129425172861202]},
    "temperature": {"n": 1, "mean": 1.0336749026740242, "std": null, "lo": null, "hi": null, "values": [1.0336749026740242]},
    "ece_temp": {"n": 1, "mean": 0.1951686630584362, "std": null, "lo": null, "hi": null, "values": [0.1951686630584362]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.1946373879622148, "std": null, "lo": null, "hi": null, "values": [0.1946373879622148]},
    "mce_temp": {"n": 1, "mean": 0.9859840297616348, "std": null, "lo": null, "hi": null, "values": [0.9859840297616348]},
    "brier_temp": {"n": 1, "mean": 0.212994521192015, "std": null, "lo": null, "hi": null, "values": [0.212994521192015]},
    "ece_platt": {"n": 1, "mean": 0.0203211708481702, "std": null, "lo": null, "hi": null, "values": [0.0203211708481702]},
    "adaptive_ece_platt": {"n": 1, "mean": 0.0299451778929004, "std": null, "lo": null, "hi": null, "values": [0.0299451778929004]},
    "mce_platt": {"n": 1, "mean": 0.6899030642125463, "std": null, "lo": null, "hi": null, "values": [0.6899030642125463]},
    "brier_platt": {"n": 1, "mean": 0.162962070978078, "std": null, "lo": null, "hi": null, "values": [0.162962070978078]}
   }},
  {"model": "resnet", "task": "ptbxl-multilabel", "averaging": "macro-label", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.6789972024991121, "std": null, "lo": null, "hi": null, "values": [0.6789972024991121]},
    "auroc": {"n": 1, "mean": 0.9112057705647688, "std": null, "lo": null, "hi": null, "values": [0.9112057705647688]},
    "precision": {"n": 1, "mean": 0.5837374092473124, "std": null, "lo": null, "hi": null, "values": [0.5837374092473124]},
    "recall": {"n": 1, "mean": 0.8558941485777506, "std": null, "lo": null, "hi": null, "values": [0.8558941485777506]},
    "exact_match_accuracy": {"n": 1, "mean": 0.4119555143651529, "std": null, "lo": null, "hi": null, "values": [0.4119555143651529]},
    "label_accuracy": {"n": 1, "mean": 0.8145505097312327, "std": null, "lo": null, "hi": null, "values": [0.8145505097312327]},
    "ece": {"n": 1, "mean": 0.1451788139279712, "std": null, "lo": null, "hi": null, "values": [0.1451788139279712]},
    "adaptive_ece": {"n": 1, "mean": 0.1448282322924315, "std": null, "lo": null, "hi": null, "values": [0.1448282322924315]},
    "mce": {"n": 1, "mean": 0.5276219622198358, "std": null, "lo": null, "hi": null, "values": [0.5276219622198358]},
    "brier": {"n": 1, "mean": 0.1290665184013458, "std": null, "lo": null, "hi": null, "values": [0.1290665184013458]},
    "temperature": {"n": 1, "mean": 1.161929319504004, "std": null, "lo": null, "hi": null, "values": [1.161929319504004]},
    "ece_temp": {"n": 1, "mean": 0.1510616660952826, "std": null, "lo": null, "hi": null, "values": [0.1510616660952826]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.1508994025917835, "std": null, "lo": null, "hi": null, "values": [0.1508994025917835]},
    "mce_temp": {"n": 1, "mean": 0.4968916033356637, "std": null, "lo": null, "hi": null, "values": [0.4968916033356637]},
    "brier_temp": {"n": 1, "mean": 0.1284086687582999, "std": null, "lo": null, "hi": null, "values": [0.1284086687582999]},
    "ece_platt": {"n": 1, "mean": 0.0217030411055077, "std": null, "lo": null, "hi": null, "values": [0.0217030411055077]},
    "adaptive_ece_platt": {"n": 1, "mean": 0.0187541275847942, "std": null, "lo": null, "hi": null, "values": [0.0187541275847942]},
    "mce_platt": {"n": 1, "mean": 0.278934993276591, "std": null, "lo": null, "hi": null, "values": [0.278934993276591]},
    "brier_platt": {"n": 1, "mean": 0.0878872265334813, "std": null, "lo": null, "hi": null, "values": [0.0878872265334813]}
   }},
  {"model": "resnet-hannun", "task": "ptbxl-multilabel", "averaging": "macro-label", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.6866251337836629, "std": null, "lo": null, "hi": null, "values": [0.6866251337836629]},
    "auroc": {"n": 1, "mean": 0.9090198897587036, "std": null, "lo": null, "hi": null, "values": [0.9090198897587036]},
    "precision": {"n": 1, "mean": 0.6005052575059564, "std": null, "lo": null, "hi": null, "values": [0.6005052575059564]},
    "recall": {"n": 1, "mean": 0.8439781449739006, "std": null, "lo": null, "hi": null, "values": [0.8439781449739006]},
    "exact_match_accuracy": {"n": 1, "mean": 0.4369786839666357, "std": null, "lo": null, "hi": null, "values": [0.4369786839666357]},
    "label_accuracy": {"n": 1, "mean": 0.8216867469879519, "std": null, "lo": null, "hi": null, "values": [0.8216867469879519]},
    "ece": {"n": 1, "mean": 0.119811119499355, "std": null, "lo": null, "hi": null, "values": [0.119811119499355]},
    "adaptive_ece": {"n": 1, "mean": 0.1207212894531081, "std": null, "lo": null, "hi": null, "values": [0.1207212894531081]},
    "mce": {"n": 1, "mean": 0.5402288570695994, "std": null, "lo": null, "hi": null, "values": [0.5402288570695994]},
    "brier": {"n": 1, "mean": 0.1204069215581235, "std": null, "lo": null, "hi": null, "values": [0.1204069215581235]},
    "temperature": {"n": 1, "mean": 1.1211007796627066, "std": null, "lo": null, "hi": null, "values": [1.1211007796627066]},
    "ece_temp": {"n": 1, "mean": 0.1243113157864579, "std": null, "lo": null, "hi": null, "values": [0.1243113157864579]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.1248913427836367, "std": null, "lo": null, "hi": null, "values": [0.1248913427836367]},
    "mce_temp": {"n": 1, "mean": 0.5013705981467567, "std": null, "lo": null, "hi": null, "values": [0.5013705981467567]},
    "brier_temp": {"n": 1, "mean": 0.1199494042078728, "std": null, "lo": null, "hi": null, "values": [0.1199494042078728]},
    "ece_platt": {"n": 1, "mean": 0.0162441240082366, "std": null, "lo": null, "hi": null, "values": [0.0162441240082366]},
    "adaptive_ece_platt": {"n": 1, "mean": 0.01724413074499, "std": null, "lo": null, "hi": null, "values": [0.01724413074499]},
    "mce_platt": {"n": 1, "mean": 0.2165813287525404, "std": null, "lo": null, "hi": null, "values": [0.2165813287525404]},
    "brier_platt": {"n": 1, "mean": 0.0879189256161275, "std": null, "lo": null, "hi": null, "values": [0.0879189256161275]}
   }},
  {"model": "resnet-wang", "task": "ptbxl-multilabel", "averaging": "macro-label", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.6765326710764995, "std": null, "lo": null, "hi": null, "values": [0.6765326710764995]},
    "auroc": {"n": 1, "mean": 0.8924301403006967, "std": null, "lo": null, "hi": null, "values": [0.8924301403006967]},
    "precision": {"n": 1, "mean": 0.6079222112272278, "std": null, "lo": null, "hi": null, "values": [0.6079222112272278]},
    "recall": {"n": 1, "mean": 0.7983140870895712, "std": null, "lo": null, "hi": null, "values": [0.7983140870895712]},
    "exact_match_accuracy": {"n": 1, "mean": 0.4578313253012048, "std": null, "lo": null, "hi": null, "values": [0.4578313253012048]},
    "label_accuracy": {"n": 1, "mean": 0.823725671918443, "std": null, "lo": null, "hi": null, "values": [0.823725671918443]},
    "ece": {"n": 1, "mean": 0.1171871255540469, "std": null, "lo": null, "hi": null, "values": [0.1171871255540469]},
    "adaptive_ece": {"n": 1, "mean": 0.1174105206923978, "std": null, "lo": null, "hi": null, "values": [0.1174105206923978]},
    "mce": {"n": 1, "mean": 0.499786562733836, "std": null, "lo": null, "hi": null, "values": [0.499786562733836]},
    "brier": {"n": 1, "mean": 0.1234788661748361, "std": null, "lo": null, "hi": null, "values": [0.1234788661748361]},
    "temperature": {"n": 1, "mean": 1.1142704621139718, "std": null, "lo": null, "hi": null, "values": [1.1142704621139718]},
    "ece_temp": {"n": 1, "mean": 0.1207310866733546, "std": null, "lo": null, "hi": null, "values": [0.1207310866733546]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.1214462319426654, "std": null, "lo": null, "hi": null, "values": [0.1214462319426654]},
    "mce_temp": {"n": 1, "mean": 0.4701932069622156, "std": null, "lo": null, "hi": null, "values": [0.4701932069622156]},
    "brier_temp": {"n": 1, "mean": 0.1236034282034026, "std": null, "lo": null, "hi": null, "values": [0.1236034282034026]},
    "ece_platt": {"n": 1, "mean": 0.0193005227112371, "std": null, "lo": null, "hi": null, "values": [0.0193005227112371]},
    "adaptive_ece_platt": {"n": 1, "mean": 0.0199469283179807, "std": null, "lo": null, "hi": null, "values": [0.0199469283179807]},
    "mce_platt": {"n": 1, "mean": 0.8053164321570585, "std": null, "lo": null, "hi": null, "values": [0.8053164321570585]},
    "brier_platt": {"n": 1, "mean": 0.0964403892772862, "std": null, "lo": null, "hi": null, "values": [0.0964403892772862]}
   }},
  {"model": "rocket", "task": "ptbxl-multilabel", "averaging": "macro-label", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.6466517398611901, "std": null, "lo": null, "hi": null, "values": [0.6466517398611901]},
    "auroc": {"n": 1, "mean": 0.8738176069360952, "std": null, "lo": null, "hi": null, "values": [0.8738176069360952]},
    "precision": {"n": 1, "mean": 0.564232398560614, "std": null, "lo": null, "hi": null, "values": [0.564232398560614]},
    "recall": {"n": 1, "mean": 0.7706417474810069, "std": null, "lo": null, "hi": null, "values": [0.7706417474810069]},
    "exact_match_accuracy": {"n": 1, "mean": 0.3943466172381835, "std": null, "lo": null, "hi": null, "values": [0.3943466172381835]},
    "label_accuracy": {"n": 1, "mean": 0.8091751621872104, "std": null, "lo": null, "hi": null, "values": [0.8091751621872104]},
    "ece": {"n": 1, "mean": 0.1326531500865266, "std": null, "lo": null, "hi": null, "values": [0.1326531500865266]},
    "adaptive_ece": {"n": 1, "mean": 0.1326941744913626, "std": null, "lo": null, "hi": null, "values": [0.1326941744913626]},
    "mce": {"n": 1, "mean": 0.5388993584192716, "std": null, "lo": null, "hi": null, "values": [0.5388993584192716]},
    "brier": {"n": 1, "mean": 0.1347224784237886, "std": null, "lo": null, "hi": null, "values": [0.1347224784237886]},
    "temperature": {"n": 1, "mean": 1.1060329571052785, "std": null, "lo": null, "hi": null, "values": [1.1060329571052785]},
    "ece_temp": {"n": 1, "mean": 0.1375933918752604, "std": null, "lo": null, "hi": null, "values": [0.1375933918752604]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.1375933918752604, "std": null, "lo": null, "hi": null, "values": [0.1375933918752604]},
    "mce_temp": {"n": 1, "mean": 0.5055180288143191, "std": null, "lo": null, "hi": null, "values": [0.5055180288143191]},
    "brier_temp": {"n": 1, "mean": 0.1345466585366513, "std": null, "lo": null, "hi": null, "values": [0.1345466585366513]},
    "ece_platt": {"n": 1, "mean": 0.0272171726509708, "std": null, "lo": null, "hi": null, "values": [0.0272171726509708]},
    "adaptive_ece_platt": {"n": 1, "mean": 0.0282148986710342, "std": null, "lo": null, "hi": null, "values": [0.0282148986710342]},
    "mce_platt": {"n": 1, "mean": 0.4420220624551919, "std": null, "lo": null, "hi": null, "values": [0.4420220624551919]},
    "brier_platt": {"n": 1, "mean": 0.1055970629560478, "std": null, "lo": null, "hi": null, "values": [0.1055970629560478]}
   }},
  {"model": "se-resnet", "task": "ptbxl-multilabel", "averaging": "macro-label", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.7036354692591633, "std": null, "lo": null, "hi": null, "values": [0.7036354692591633]},
    "auroc": {"n": 1, "mean": 0.912706296322952, "std": null, "lo": null, "hi": null, "values": [0.912706296322952]},
    "precision": {"n": 1, "mean": 0.6238459371312126, "std": null, "lo": null, "hi": null, "values": [0.6238459371312126]},
    "recall": {"n": 1, "mean": 0.8391595521282742, "std": null, "lo": null, "hi": null, "values": [0.8391595521282742]},
    "exact_match_accuracy": {"n": 1, "mean": 0.4768303985171455, "std": null, "lo": null, "hi": null, "values": [0.4768303985171455]},
    "label_accuracy": {"n": 1, "mean": 0.8388322520852641, "std": null, "lo": null, "hi": null, "values": [0.8388322520852641]},
    "ece": {"n": 1, "mean": 0.1100364312008399, "std": null, "lo": null, "hi": null, "values": [0.1100364312008399]},
    "adaptive_ece": {"n": 1, "mean": 0.1091763300353534, "std": null, "lo": null, "hi": null, "values": [0.1091763300353534]},
    "mce": {"n": 1, "mean": 0.5330017437768537, "std": null, "lo": null, "hi": null, "values": [0.5330017437768537]},
    "brier": {"n": 1, "mean": 0.1136192387671581, "std": null, "lo": null, "hi": null, "values": [0.1136192387671581]},
    "temperature": {"n": 1, "mean": 1.086894183510598, "std": null, "lo": null, "hi": null, "values": [1.086894183510598]},
    "ece_temp": {"n": 1, "mean": 0.1123662179729811, "std": null, "lo": null, "hi": null, "values": [0.1123662179729811]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.1123085300565645, "std": null, "lo": null, "hi": null, "values": [0.1123085300565645]},
    "mce_temp": {"n": 1, "mean": 0.5142057388329814, "std": null, "lo": null, "hi": null, "values": [0.5142057388329814]},
    "brier_temp": {"n": 1, "mean": 0.1134227303333316, "std": null, "lo": null, "hi": null, "values": [0.1134227303333316]},
    "ece_platt": {"n": 1, "mean": 0.0183230236011649, "std": null, "lo": null, "hi": null, "values": [0.0183230236011649]},
    "adaptive_ece_platt": {"n": 1, "mean": 0.01636703343388, "std": null, "lo": null, "hi": null, "values": [0.01636703343388]},
    "mce_platt": {"n": 1, "mean": 0.177616292441967, "std": null, "lo": null, "hi": null, "values": [0.177616292441967]},
    "brier_platt": {"n": 1, "mean": 0.0866122408199547, "std": null, "lo": null, "hi": null, "values": [0.0866122408199547]}
   }},
  {"model": "simple-cnn", "task": "ptbxl-multilabel", "averaging": "macro-label", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.6772147985217829, "std": null, "lo": null, "hi": null, "values": [0.6772147985217829]},
    "auroc": {"n": 1, "mean": 0.893949272491704, "std": null, "lo": null, "hi": null, "values": [0.893949272491704]},
    "precision": {"n": 1, "mean": 0.604535239496988, "std": null, "lo": null, "hi": null, "values": [0.604535239496988]},
    "recall": {"n": 1, "mean": 0.8171098979193806, "std": null, "lo": null, "hi": null, "values": [0.8171098979193806]},
    "exact_match_accuracy": {"n": 1, "mean": 0.4267840593141798, "std": null, "lo": null, "hi": null, "values": [0.4267840593141798]},
    "label_accuracy": {"n": 1, "mean": 0.8171455050973123, "std": null, "lo": null, "hi": null, "values": [0.8171455050973123]},
    "ece": {"n": 1, "mean": 0.129830318578066, "std": null, "lo": null, "hi": null, "values": [0.129830318578066]},
    "adaptive_ece": {"n": 1, "mean": 0.1294421056023748, "std": null, "lo": null, "hi": null, "values": [0.1294421056023748]},
    "mce": {"n": 1, "mean": 0.5662345646358118, "std": null, "lo": null, "hi": null, "values": [0.5662345646358118]},
    "brier": {"n": 1, "mean": 0.1300354734918702, "std": null, "lo": null, "hi": null, "values": [0.1300354734918702]},
    "temperature": {"n": 1, "mean": 1.1349028143674105, "std": null, "lo": null, "hi": null, "values": [1.1349028143674105]},
    "ece_temp": {"n": 1, "mean": 0.1349506908383227, "std": null, "lo": null, "hi": null, "values": [0.1349506908383227]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.1348406755916859, "std": null, "lo": null, "hi": null, "values": [0.1348406755916859]},
    "mce_temp": {"n": 1, "mean": 0.520979577890745, "std": null, "lo": null, "hi": null, "values": [0.520979577890745]},
    "brier_temp": {"n": 1, "mean": 0.1299310289101949, "std": null, "lo": null, "hi": null, "values": [0.1299310289101949]},
    "ece_platt": {"n": 1, "mean": 0.0176822529744991, "std": null, "lo": null, "hi": null, "values": [0.0176822529744991]},
    "adaptive_ece_platt": {"n": 1, "mean": 0.0177533844311385, "std": null, "lo": null, "hi": null, "values": [0.0177533844311385]},
    "mce_platt": {"n": 1, "mean": 0.3639735634121672, "std": null, "lo": null, "hi": null, "values": [0.3639735634121672]},
    "brier_platt": {"n": 1, "mean": 0.0957014503304969, "std": null, "lo": null, "hi": null, "values": [0.0957014503304969]}
   }},
  {"model": "spectrogram-cnn", "task": "ptbxl-multilabel", "averaging": "macro-label", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.6597804836255261, "std": null, "lo": null, "hi": null, "values": [0.6597804836255261]},
    "auroc": {"n": 1, "mean": 0.8933837832850691, "std": null, "lo": null, "hi": null, "values": [0.8933837832850691]},
    "precision": {"n": 1, "mean": 0.559931450282395, "std": null, "lo": null, "hi": null, "values": [0.559931450282395]},
    "recall": {"n": 1, "mean": 0.832595337954228, "std": null, "lo": null, "hi": null, "values": [0.832595337954228]},
    "exact_match_accuracy": {"n": 1, "mean": 0.4050046339202965, "std": null, "lo": null, "hi": null, "values": [0.4050046339202965]},
    "label_accuracy": {"n": 1, "mean": 0.8029657089898053, "std": null, "lo": null, "hi": null, "values": [0.8029657089898053]},
    "ece": {"n": 1, "mean": 0.1413944977095768, "std": null, "lo": null, "hi": null, "values": [0.1413944977095768]},
    "adaptive_ece": {"n": 1, "mean": 0.1408891566771932, "std": null, "lo": null, "hi": null, "values": [0.1408891566771932]},
    "mce": {"n": 1, "mean": 0.5033215456887294, "std": null, "lo": null, "hi": null, "values": [0.5033215456887294]},
    "brier": {"n": 1, "mean": 0.1355286813251555, "std": null, "lo": null, "hi": null, "values": [0.1355286813251555]},
    "temperature": {"n": 1, "mean": 1.0954557457627332, "std": null, "lo": null, "hi": null, "values": [1.0954557457627332]},
    "ece_temp": {"n": 1, "mean": 0.1454672370574095, "std": null, "lo": null, "hi": null, "values": [0.1454672370574095]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.1437276523110646, "std": null, "lo": null, "hi": null, "values": [0.1437276523110646]},
    "mce_temp": {"n": 1, "mean": 0.4897803789676787, "std": null, "lo": null, "hi": null, "values": [0.4897803789676787]},
    "brier_temp": {"n": 1, "mean": 0.135123084754264, "std": null, "lo": null, "hi": null, "values": [0.135123084754264]},
    "ece_platt": {"n": 1, "mean": 0.0256724150686066, "std": null, "lo": null, "hi": null, "values": [0.0256724150686066]},
    "adaptive_ece_platt": {"n": 1, "mean": 0.0215466680859352, "std": null, "lo": null, "hi": null, "values": [0.0215466680859352]},
    "mce_platt": {"n": 1, "mean": 0.2625360685308441, "std": null, "lo": null, "hi": null, "values": [0.2625360685308441]},
    "brier_platt": {"n": 1, "mean": 0.0977930959681375, "std": null, "lo": null, "hi": null, "values": [0.0977930959681375]}
   }},
  {"model": "tcn", "task": "ptbxl-multilabel", "averaging": "macro-label", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.7016444969111284, "std": null, "lo": null, "hi": null, "values": [0.7016444969111284]},
    "auroc": {"n": 1, "mean": 0.9100086297727114, "std": null, "lo": null, "hi": null, "values": [0.9100086297727114]},
    "precision": {"n": 1, "mean": 0.6212201555668156, "std": null, "lo": null, "hi": null, "values": [0.6212201555668156]},
    "recall": {"n": 1, "mean": 0.8203410353958628, "std": null, "lo": null, "hi": null, "values": [0.8203410353958628]},
    "exact_match_accuracy": {"n": 1, "mean": 0.4856348470806302, "std": null, "lo": null, "hi": null, "values": [0.4856348470806302]},
    "label_accuracy": {"n": 1, "mean": 0.8435588507877665, "std": null, "lo": null, "hi": null, "values": [0.8435588507877665]},
    "ece": {"n": 1, "mean": 0.1037826201657896, "std": null, "lo": null, "hi": null, "values": [0.1037826201657896]},
    "adaptive_ece": {"n": 1, "mean": 0.1046893235374495, "std": null, "lo": null, "hi": null, "values": [0.1046893235374495]},
    "mce": {"n": 1, "mean": 0.4506663003664338, "std": null, "lo": null, "hi": null, "values": [0.4506663003664338]},
    "brier": {"n": 1, "mean": 0.1097733508988451, "std": null, "lo": null, "hi": null, "values": [0.1097733508988451]},
    "temperature": {"n": 1, "mean": 1.0790570394941987, "std": null, "lo": null, "hi": null, "values": [1.0790570394941987]},
    "ece_temp": {"n": 1, "mean": 0.1066075127383397, "std": null, "lo": null, "hi": null, "values": [0.1066075127383397]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.106894161929101, "std": null, "lo": null, "hi": null, "values": [0.106894161929101]},
    "mce_temp": {"n": 1, "mean": 0.3921011202599676, "std": null, "lo": null, "hi": null, "values": [0.3921011202599676]},
    "brier_temp": {"n": 1, "mean": 0.1095816173711432, "std": null, "lo": null, "hi": null, "values": [0.1095816173711432]},
    "ece_platt": {"n": 1, "mean": 0.0181371011683854, "std": null, "lo": null, "hi": null, "values": [0.0181371011683854]},
    "adaptive_ece_platt": {"n": 1, "mean": 0.0145115656694394, "std": null, "lo": null, "hi": null, "values": [0.0145115656694394]},
    "mce_platt": {"n": 1, "mean": 0.1780047929054293, "std": null, "lo": null, "hi": null, "values": [0.1780047929054293]},
    "brier_platt": {"n": 1, "mean": 0.0883480220296756, "std": null, "lo": null, "hi": null, "values": [0.0883480220296756]}
   }},
  {"model": "transformer", "task": "ptbxl-multilabel", "averaging": "macro-label", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.5852773129263105, "std": null, "lo": null, "hi": null, "values": [0.5852773129263105]},
    "auroc": {"n": 1, "mean": 0.8250017994652806, "std": null, "lo": null, "hi": null, "values": [0.8250017994652806]},
    "precision": {"n": 1, "mean": 0.5016101019387065, "std": null, "lo": null, "hi": null, "values": [0.5016101019387065]},
    "recall": {"n": 1, "mean": 0.723444402791513, "std": null, "lo": null, "hi": null, "values": [0.723444402791513]},
    "exact_match_accuracy": {"n": 1, "mean": 0.3948100092678406, "std": null, "lo": null, "hi": null, "values": [0.3948100092678406]},
    "label_accuracy": {"n": 1, "mean": 0.7619091751621873, "std": null, "lo": null, "hi": null, "values": [0.7619091751621873]},
    "ece": {"n": 1, "mean": 0.1459825413361305, "std": null, "lo": null, "hi": null, "values": [0.1459825413361305]},
    "adaptive_ece": {"n": 1, "mean": 0.1457589234576047, "std": null, "lo": null, "hi": null, "values": [0.1457589234576047]},
    "mce": {"n": 1, "mean": 0.4175006799838122, "std": null, "lo": null, "hi": null, "values": [0.4175006799838122]},
    "brier": {"n": 1, "mean": 0.1559294657729887, "std": null, "lo": null, "hi": null, "values": [0.1559294657729887]},
    "temperature": {"n": 1, "mean": 0.9261313461931406, "std": null, "lo": null, "hi": null, "values": [0.9261313461931406]},
    "ece_temp": {"n": 1, "mean": 0.1425562539687667, "std": null, "lo": null, "hi": null, "values": [0.1425562539687667]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.1423830418839544, "std": null, "lo": null, "hi": null, "values": [0.1423830418839544]},
    "mce_temp": {"n": 1, "mean": 0.4593408150029986, "std": null, "lo": null, "hi": null, "values": [0.4593408150029986]},
    "brier_temp": {"n": 1, "mean": 0.156007914429961, "std": null, "lo": null, "hi": null, "values": [0.156007914429961]},
    "ece_platt": {"n": 1, "mean": 0.0281326222655352, "std": null, "lo": null, "hi": null, "values": [0.0281326222655352]},
    "adaptive_ece_platt": {"n": 1, "mean": 0.0282561581244359, "std": null, "lo": null, "hi": null, "values": [0.0282561581244359]},
    "mce_platt": {"n": 1, "mean": 0.3429312405685789, "std": null, "lo": null, "hi": null, "values": [0.3429312405685789]},
    "brier_platt": {"n": 1, "mean": 0.1253414230335514, "std": null, "lo": null, "hi": null, "values": [0.1253414230335514]}
   }},
  {"model": "xresnet", "task": "ptbxl-multilabel", "averaging": "macro-label", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.6967876801737842, "std": null, "lo": null, "hi": null, "values": [0.6967876801737842]},
    "auroc": {"n": 1, "mean": 0.9097583329057974, "std": null, "lo": null, "hi": null, "values": [0.9097583329057974]},
    "precision": {"n": 1, "mean": 0.6055984748666591, "std": null, "lo": null, "hi": null, "values": [0.6055984748666591]},
    "recall": {"n": 1, "mean": 0.8464582530214901, "std": null, "lo": null, "hi": null, "values": [0.8464582530214901]},
    "exact_match_accuracy": {"n": 1, "mean": 0.4620018535681186, "std": null, "lo": null, "hi": null, "values": [0.4620018535681186]},
    "label_accuracy": {"n": 1, "mean": 0.8337349397590361, "std": null, "lo": null, "hi": null, "values": [0.8337349397590361]},
    "ece": {"n": 1, "mean": 0.1228584012965529, "std": null, "lo": null, "hi": null, "values": [0.1228584012965529]},
    "adaptive_ece": {"n": 1, "mean": 0.1230174967871821, "std": null, "lo": null, "hi": null, "values": [0.1230174967871821]},
    "mce": {"n": 1, "mean": 0.4418094052987941, "std": null, "lo": null, "hi": null, "values": [0.4418094052987941]},
    "brier": {"n": 1, "mean": 0.116658436853299, "std": null, "lo": null, "hi": null, "values": [0.116658436853299]},
    "temperature": {"n": 1, "mean": 0.9742281218518292, "std": null, "lo": null, "hi": null, "values": [0.9742281218518292]},
    "ece_temp": {"n": 1, "mean": 0.1218474320857711, "std": null, "lo": null, "hi": null, "values": [0.1218474320857711]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.121974187577716, "std": null, "lo": null, "hi": null, "values": [0.121974187577716]},
    "mce_temp": {"n": 1, "mean": 0.4440814241018741, "std": null, "lo": null, "hi": null, "values": [0.4440814241018741]},
    "brier_temp": {"n": 1, "mean": 0.1166189186460035, "std": null, "lo": null, "hi": null, "values": [0.1166189186460035]},
    "ece_platt": {"n": 1, "mean": 0.0176574499468985, "std": null, "lo": null, "hi": null, "values": [0.0176574499468985]},
    "adaptive_ece_platt": {"n": 1, "mean": 0.0171978859435512, "std": null, "lo": null, "hi": null, "values": [0.0171978859435512]},
    "mce_platt": {"n": 1, "mean": 0.1621691978851797, "std": null, "lo": null, "hi": null, "values": [0.1621691978851797]},
    "brier_platt": {"n": 1, "mean": 0.0883251866705623, "std": null, "lo": null, "hi": null, "values": [0.0883251866705623]}
   }},
  {"model": "xresnet1d-benchmark", "task": "ptbxl-multilabel", "averaging": "macro-label", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.6858804981086027, "std": null, "lo": null, "hi": null, "values": [0.6858804981086027]},
    "auroc": {"n": 1, "mean": 0.9074839968697136, "std": null, "lo": null, "hi": null, "values": [0.9074839968697136]},
    "precision": {"n": 1, "mean": 0.5965386827874478, "std": null, "lo": null, "hi": null, "values": [0.5965386827874478]},
    "recall": {"n": 1, "mean": 0.8443316955314953, "std": null, "lo": null, "hi": null, "values": [0.8443316955314953]},
    "exact_match_accuracy": {"n": 1, "mean": 0.4314179796107507, "std": null, "lo": null, "hi": null, "values": [0.4314179796107507]},
    "label_accuracy": {"n": 1, "mean": 0.8223354958294717, "std": null, "lo": null, "hi": null, "values": [0.8223354958294717]},
    "ece": {"n": 1, "mean": 0.1289829151061886, "std": null, "lo": null, "hi": null, "values": [0.1289829151061886]},
    "adaptive_ece": {"n": 1, "mean": 0.1283568785900472, "std": null, "lo": null, "hi": null, "values": [0.1283568785900472]},
    "mce": {"n": 1, "mean": 0.600873589515686, "std": null, "lo": null, "hi": null, "values": [0.600873589515686]},
    "brier": {"n": 1, "mean": 0.1260612779424896, "std": null, "lo": null, "hi": null, "values": [0.1260612779424896]},
    "temperature": {"n": 1, "mean": 1.2145359538473368, "std": null, "lo": null, "hi": null, "values": [1.2145359538473368]},
    "ece_temp": {"n": 1, "mean": 0.1361975090551318, "std": null, "lo": null, "hi": null, "values": [0.1361975090551318]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.1361656986287548, "std": null, "lo": null, "hi": null, "values": [0.1361656986287548]},
    "mce_temp": {"n": 1, "mean": 0.5748700830960352, "std": null, "lo": null, "hi": null, "values": [0.5748700830960352]},
    "brier_temp": {"n": 1, "mean": 0.1252653586392059, "std": null, "lo": null, "hi": null, "values": [0.1252653586392059]},
    "ece_platt": {"n": 1, "mean": 0.0207766771341708, "std": null, "lo": null, "hi": null, "values": [0.0207766771341708]},
    "adaptive_ece_platt": {"n": 1, "mean": 0.0205089695886622, "std": null, "lo": null, "hi": null, "values": [0.0205089695886622]},
    "mce_platt": {"n": 1, "mean": 0.169680296967959, "std": null, "lo": null, "hi": null, "values": [0.169680296967959]},
    "brier_platt": {"n": 1, "mean": 0.0901525850238901, "std": null, "lo": null, "hi": null, "values": [0.0901525850238901]}
   }},
  {"model": "xresnet1d101", "task": "ptbxl-multilabel", "averaging": "macro-label", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.6993682462678301, "std": null, "lo": null, "hi": null, "values": [0.6993682462678301]},
    "auroc": {"n": 1, "mean": 0.9092362153780168, "std": null, "lo": null, "hi": null, "values": [0.9092362153780168]},
    "precision": {"n": 1, "mean": 0.632670283094406, "std": null, "lo": null, "hi": null, "values": [0.632670283094406]},
    "recall": {"n": 1, "mean": 0.8174494994698807, "std": null, "lo": null, "hi": null, "values": [0.8174494994698807]},
    "exact_match_accuracy": {"n": 1, "mean": 0.4777571825764596, "std": null, "lo": null, "hi": null, "values": [0.4777571825764596]},
    "label_accuracy": {"n": 1, "mean": 0.8395736793327154, "std": null, "lo": null, "hi": null, "values": [0.8395736793327154]},
    "ece": {"n": 1, "mean": 0.1008510767362879, "std": null, "lo": null, "hi": null, "values": [0.1008510767362879]},
    "adaptive_ece": {"n": 1, "mean": 0.1019183105046235, "std": null, "lo": null, "hi": null, "values": [0.1019183105046235]},
    "mce": {"n": 1, "mean": 0.6253744828146557, "std": null, "lo": null, "hi": null, "values": [0.6253744828146557]},
    "brier": {"n": 1, "mean": 0.1151208018329707, "std": null, "lo": null, "hi": null, "values": [0.1151208018329707]},
    "temperature": {"n": 1, "mean": 1.2476260673486412, "std": null, "lo": null, "hi": null, "values": [1.2476260673486412]},
    "ece_temp": {"n": 1, "mean": 0.1092056733814738, "std": null, "lo": null, "hi": null, "values": [0.1092056733814738]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.1080136982696383, "std": null, "lo": null, "hi": null, "values": [0.1080136982696383]},
    "mce_temp": {"n": 1, "mean": 0.533627299179678, "std": null, "lo": null, "hi": null, "values": [0.533627299179678]},
    "brier_temp": {"n": 1, "mean": 0.1140987078175659, "std": null, "lo": null, "hi": null, "values": [0.1140987078175659]},
    "ece_platt": {"n": 1, "mean": 0.0204197728691533, "std": null, "lo": null, "hi": null, "values": [0.0204197728691533]},
    "adaptive_ece_platt": {"n": 1, "mean": 0.0213086324584984, "std": null, "lo": null, "hi": null, "values": [0.0213086324584984]},
    "mce_platt": {"n": 1, "mean": 0.2316171788512259, "std": null, "lo": null, "hi": null, "values": [0.2316171788512259]},
    "brier_platt": {"n": 1, "mean": 0.0884514137181401, "std": null, "lo": null, "hi": null, "values": [0.0884514137181401]}
   }},
  {"model": "convnext-v2", "task": "physionet2017-4class", "averaging": "macro", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.6555994847045018, "std": null, "lo": null, "hi": null, "values": [0.6555994847045018]},
    "auroc": {"n": 1, "mean": 0.9147553641429238, "std": null, "lo": null, "hi": null, "values": [0.9147553641429238]},
    "accuracy": {"n": 1, "mean": 0.7859270290394639, "std": null, "lo": null, "hi": null, "values": [0.7859270290394639]},
    "mcc": {"n": 1, "mean": 0.610178933460446, "std": null, "lo": null, "hi": null, "values": [0.610178933460446]},
    "precision": {"n": 1, "mean": 0.6533000906867366, "std": null, "lo": null, "hi": null, "values": [0.6533000906867366]},
    "recall": {"n": 1, "mean": 0.6832795986286528, "std": null, "lo": null, "hi": null, "values": [0.6832795986286528]},
    "ece": {"n": 1, "mean": 0.031256790058611, "std": null, "lo": null, "hi": null, "values": [0.031256790058611]},
    "adaptive_ece": {"n": 1, "mean": 0.0287309281766547, "std": null, "lo": null, "hi": null, "values": [0.0287309281766547]},
    "mce": {"n": 1, "mean": 0.1477002231847672, "std": null, "lo": null, "hi": null, "values": [0.1477002231847672]},
    "brier": {"n": 1, "mean": 0.3010052021501918, "std": null, "lo": null, "hi": null, "values": [0.3010052021501918]},
    "temperature": {"n": 1, "mean": 1.070162870713673, "std": null, "lo": null, "hi": null, "values": [1.070162870713673]},
    "ece_temp": {"n": 1, "mean": 0.0302881958894151, "std": null, "lo": null, "hi": null, "values": [0.0302881958894151]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0283824958572692, "std": null, "lo": null, "hi": null, "values": [0.0283824958572692]},
    "mce_temp": {"n": 1, "mean": 0.1066859942556814, "std": null, "lo": null, "hi": null, "values": [0.1066859942556814]},
    "brier_temp": {"n": 1, "mean": 0.3006564888527787, "std": null, "lo": null, "hi": null, "values": [0.3006564888527787]}
   }},
  {"model": "crnn", "task": "physionet2017-4class", "averaging": "macro", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.6706077339156019, "std": null, "lo": null, "hi": null, "values": [0.6706077339156019]},
    "auroc": {"n": 1, "mean": 0.9168257003302364, "std": null, "lo": null, "hi": null, "values": [0.9168257003302364]},
    "accuracy": {"n": 1, "mean": 0.8153387937453462, "std": null, "lo": null, "hi": null, "values": [0.8153387937453462]},
    "mcc": {"n": 1, "mean": 0.6608654824498493, "std": null, "lo": null, "hi": null, "values": [0.6608654824498493]},
    "precision": {"n": 1, "mean": 0.7203475234559076, "std": null, "lo": null, "hi": null, "values": [0.7203475234559076]},
    "recall": {"n": 1, "mean": 0.6507272353427562, "std": null, "lo": null, "hi": null, "values": [0.6507272353427562]},
    "ece": {"n": 1, "mean": 0.0210318075061643, "std": null, "lo": null, "hi": null, "values": [0.0210318075061643]},
    "adaptive_ece": {"n": 1, "mean": 0.0297842543802417, "std": null, "lo": null, "hi": null, "values": [0.0297842543802417]},
    "mce": {"n": 1, "mean": 0.0651846826076507, "std": null, "lo": null, "hi": null, "values": [0.0651846826076507]},
    "brier": {"n": 1, "mean": 0.2819731457475254, "std": null, "lo": null, "hi": null, "values": [0.2819731457475254]},
    "temperature": {"n": 1, "mean": 1.1489104599273452, "std": null, "lo": null, "hi": null, "values": [1.1489104599273452]},
    "ece_temp": {"n": 1, "mean": 0.0223376681143493, "std": null, "lo": null, "hi": null, "values": [0.0223376681143493]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0327185853699541, "std": null, "lo": null, "hi": null, "values": [0.0327185853699541]},
    "mce_temp": {"n": 1, "mean": 0.0877878276613954, "std": null, "lo": null, "hi": null, "values": [0.0877878276613954]},
    "brier_temp": {"n": 1, "mean": 0.2820029957878965, "std": null, "lo": null, "hi": null, "values": [0.2820029957878965]}
   }},
  {"model": "deep-res-cnn", "task": "physionet2017-4class", "averaging": "macro", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.5654916120063567, "std": null, "lo": null, "hi": null, "values": [0.5654916120063567]},
    "auroc": {"n": 1, "mean": 0.8713501879274483, "std": null, "lo": null, "hi": null, "values": [0.8713501879274483]},
    "accuracy": {"n": 1, "mean": 0.7204020848845868, "std": null, "lo": null, "hi": null, "values": [0.7204020848845868]},
    "mcc": {"n": 1, "mean": 0.4732359462071744, "std": null, "lo": null, "hi": null, "values": [0.4732359462071744]},
    "precision": {"n": 1, "mean": 0.6203923524892428, "std": null, "lo": null, "hi": null, "values": [0.6203923524892428]},
    "recall": {"n": 1, "mean": 0.5438872605154517, "std": null, "lo": null, "hi": null, "values": [0.5438872605154517]},
    "ece": {"n": 1, "mean": 0.0301642380738027, "std": null, "lo": null, "hi": null, "values": [0.0301642380738027]},
    "adaptive_ece": {"n": 1, "mean": 0.0388867287984643, "std": null, "lo": null, "hi": null, "values": [0.0388867287984643]},
    "mce": {"n": 1, "mean": 0.1746101602911949, "std": null, "lo": null, "hi": null, "values": [0.1746101602911949]},
    "brier": {"n": 1, "mean": 0.3914688107431487, "std": null, "lo": null, "hi": null, "values": [0.3914688107431487]},
    "temperature": {"n": 1, "mean": 1.111641315776561, "std": null, "lo": null, "hi": null, "values": [1.111641315776561]},
    "ece_temp": {"n": 1, "mean": 0.0356085294807231, "std": null, "lo": null, "hi": null, "values": [0.0356085294807231]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0387133843047731, "std": null, "lo": null, "hi": null, "values": [0.0387133843047731]},
    "mce_temp": {"n": 1, "mean": 0.1212321796927806, "std": null, "lo": null, "hi": null, "values": [0.1212321796927806]},
    "brier_temp": {"n": 1, "mean": 0.3914707295383152, "std": null, "lo": null, "hi": null, "values": [0.3914707295383152]}
   }},
  {"model": "dualnet", "task": "physionet2017-4class", "averaging": "macro", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.6253037333612782, "std": null, "lo": null, "hi": null, "values": [0.6253037333612782]},
    "auroc": {"n": 1, "mean": 0.9014205620502692, "std": null, "lo": null, "hi": null, "values": [0.9014205620502692]},
    "accuracy": {"n": 1, "mean": 0.7810871183916605, "std": null, "lo": null, "hi": null, "values": [0.7810871183916605]},
    "mcc": {"n": 1, "mean": 0.5964109353894983, "std": null, "lo": null, "hi": null, "values": [0.5964109353894983]},
    "precision": {"n": 1, "mean": 0.6509474124629416, "std": null, "lo": null, "hi": null, "values": [0.6509474124629416]},
    "recall": {"n": 1, "mean": 0.618258465386385, "std": null, "lo": null, "hi": null, "values": [0.618258465386385]},
    "ece": {"n": 1, "mean": 0.0354966112285608, "std": null, "lo": null, "hi": null, "values": [0.0354966112285608]},
    "adaptive_ece": {"n": 1, "mean": 0.0381282996039557, "std": null, "lo": null, "hi": null, "values": [0.0381282996039557]},
    "mce": {"n": 1, "mean": 0.1260777659275952, "std": null, "lo": null, "hi": null, "values": [0.1260777659275952]},
    "brier": {"n": 1, "mean": 0.3120819188246668, "std": null, "lo": null, "hi": null, "values": [0.3120819188246668]},
    "temperature": {"n": 1, "mean": 1.296101320977692, "std": null, "lo": null, "hi": null, "values": [1.296101320977692]},
    "ece_temp": {"n": 1, "mean": 0.0321057577972605, "std": null, "lo": null, "hi": null, "values": [0.0321057577972605]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0335995376351768, "std": null, "lo": null, "hi": null, "values": [0.0335995376351768]},
    "mce_temp": {"n": 1, "mean": 0.2257080599264623, "std": null, "lo": null, "hi": null, "values": [0.2257080599264623]},
    "brier_temp": {"n": 1, "mean": 0.3124301023295301, "std": null, "lo": null, "hi": null, "values": [0.3124301023295301]}
   }},
  {"model": "fcn-wang", "task": "physionet2017-4class", "averaging": "macro", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.3816717847941021, "std": null, "lo": null, "hi": null, "values": [0.3816717847941021]},
    "auroc": {"n": 1, "mean": 0.7584205150734864, "std": null, "lo": null, "hi": null, "values": [0.7584205150734864]},
    "accuracy": {"n": 1, "mean": 0.6299329858525688, "std": null, "lo": null, "hi": null, "values": [0.6299329858525688]},
    "mcc": {"n": 1, "mean": 0.2435568014979326, "std": null, "lo": null, "hi": null, "values": [0.2435568014979326]},
    "precision": {"n": 1, "mean": 0.5372505373154641, "std": null, "lo": null, "hi": null, "values": [0.5372505373154641]},
    "recall": {"n": 1, "mean": 0.3605615318811714, "std": null, "lo": null, "hi": null, "values": [0.3605615318811714]},
    "ece": {"n": 1, "mean": 0.0264657733494545, "std": null, "lo": null, "hi": null, "values": [0.0264657733494545]},
    "adaptive_ece": {"n": 1, "mean": 0.0288019515268854, "std": null, "lo": null, "hi": null, "values": [0.0288019515268854]},
    "mce": {"n": 1, "mean": 0.3319614827632904, "std": null, "lo": null, "hi": null, "values": [0.3319614827632904]},
    "brier": {"n": 1, "mean": 0.4899937891437532, "std": null, "lo": null, "hi": null, "values": [0.4899937891437532]},
    "temperature": {"n": 1, "mean": 0.9417061295969216, "std": null, "lo": null, "hi": null, "values": [0.9417061295969216]},
    "ece_temp": {"n": 1, "mean": 0.0238089808599165, "std": null, "lo": null, "hi": null, "values": [0.0238089808599165]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0312272549829853, "std": null, "lo": null, "hi": null, "values": [0.0312272549829853]},
    "mce_temp": {"n": 1, "mean": 0.2294241354886783, "std": null, "lo": null, "hi": null, "values": [0.2294241354886783]},
    "brier_temp": {"n": 1, "mean": 0.489871084170868, "std": null, "lo": null, "hi": null, "values": [0.489871084170868]}
   }},
  {"model": "gru", "task": "physionet2017-4class", "averaging": "macro", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.4077330032135488, "std": null, "lo": null, "hi": null, "values": [0.4077330032135488]},
    "auroc": {"n": 1, "mean": 0.7907008530960911, "std": null, "lo": null, "hi": null, "values": [0.7907008530960911]},
    "accuracy": {"n": 1, "mean": 0.6008935219657483, "std": null, "lo": null, "hi": null, "values": [0.6008935219657483]},
    "mcc": {"n": 1, "mean": 0.2811914763605249, "std": null, "lo": null, "hi": null, "values": [0.2811914763605249]},
    "precision": {"n": 1, "mean": 0.4408941928974338, "std": null, "lo": null, "hi": null, "values": [0.4408941928974338]},
    "recall": {"n": 1, "mean": 0.4977906223581424, "std": null, "lo": null, "hi": null, "values": [0.4977906223581424]},
    "ece": {"n": 1, "mean": 0.0486963289226069, "std": null, "lo": null, "hi": null, "values": [0.0486963289226069]},
    "adaptive_ece": {"n": 1, "mean": 0.0496271368024603, "std": null, "lo": null, "hi": null, "values": [0.0496271368024603]},
    "mce": {"n": 1, "mean": 0.0810237714799784, "std": null, "lo": null, "hi": null, "values": [0.0810237714799784]},
    "brier": {"n": 1, "mean": 0.5170890526320143, "std": null, "lo": null, "hi": null, "values": [0.5170890526320143]},
    "temperature": {"n": 1, "mean": 1.1414497296207002, "std": null, "lo": null, "hi": null, "values": [1.1414497296207002]},
    "ece_temp": {"n": 1, "mean": 0.0163513143648266, "std": null, "lo": null, "hi": null, "values": [0.0163513143648266]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0286573761156948, "std": null, "lo": null, "hi": null, "values": [0.0286573761156948]},
    "mce_temp": {"n": 1, "mean": 0.0537042211947289, "std": null, "lo": null, "hi": null, "values": [0.0537042211947289]},
    "brier_temp": {"n": 1, "mean": 0.5133834220175506, "std": null, "lo": null, "hi": null, "values": [0.5133834220175506]}
   }},
  {"model": "inception-time", "task": "physionet2017-4class", "averaging": "macro", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.5815102363454419, "std": null, "lo": null, "hi": null, "values": [0.5815102363454419]},
    "auroc": {"n": 1, "mean": 0.876929444746603, "std": null, "lo": null, "hi": null, "values": [0.876929444746603]},
    "accuracy": {"n": 1, "mean": 0.7475800446760983, "std": null, "lo": null, "hi": null, "values": [0.7475800446760983]},
    "mcc": {"n": 1, "mean": 0.5213636423388798, "std": null, "lo": null, "hi": null, "values": [0.5213636423388798]},
    "precision": {"n": 1, "mean": 0.6712461330332287, "std": null, "lo": null, "hi": null, "values": [0.6712461330332287]},
    "recall": {"n": 1, "mean": 0.5426747024974595, "std": null, "lo": null, "hi": null, "values": [0.5426747024974595]},
    "ece": {"n": 1, "mean": 0.0226573442689358, "std": null, "lo": null, "hi": null, "values": [0.0226573442689358]},
    "adaptive_ece": {"n": 1, "mean": 0.0343333921344619, "std": null, "lo": null, "hi": null, "values": [0.0343333921344619]},
    "mce": {"n": 1, "mean": 0.3109707534313202, "std": null, "lo": null, "hi": null, "values": [0.3109707534313202]},
    "brier": {"n": 1, "mean": 0.3702789230493597, "std": null, "lo": null, "hi": null, "values": [0.3702789230493597]},
    "temperature": {"n": 1, "mean": 1.1544400036348832, "std": null, "lo": null, "hi": null, "values": [1.1544400036348832]},
    "ece_temp": {"n": 1, "mean": 0.0457885195912579, "std": null, "lo": null, "hi": null, "values": [0.0457885195912579]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0492400476191923, "std": null, "lo": null, "hi": null, "values": [0.0492400476191923]},
    "mce_temp": {"n": 1, "mean": 0.1130711741198173, "std": null, "lo": null, "hi": null, "values": [0.1130711741198173]},
    "brier_temp": {"n": 1, "mean": 0.3714897615664189, "std": null, "lo": null, "hi": null, "values": [0.3714897615664189]}
   }},
  {"model": "kanres", "task": "physionet2017-4class", "averaging": "macro", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.4525648630834221, "std": null, "lo": null, "hi": null, "values": [0.4525648630834221]},
    "auroc": {"n": 1, "mean": 0.7948740612477376, "std": null, "lo": null, "hi": null, "values": [0.7948740612477376]},
    "accuracy": {"n": 1, "mean": 0.6548771407297096, "std": null, "lo": null, "hi": null, "values": [0.6548771407297096]},
    "mcc": {"n": 1, "mean": 0.3173504245227462, "std": null, "lo": null, "hi": null, "values": [0.3173504245227462]},
    "precision": {"n": 1, "mean": 0.5555437067396064, "std": null, "lo": null, "hi": null, "values": [0.5555437067396064]},
    "recall": {"n": 1, "mean": 0.4277526859343807, "std": null, "lo": null, "hi": null, "values": [0.4277526859343807]},
    "ece": {"n": 1, "mean": 0.0332866057653704, "std": null, "lo": null, "hi": null, "values": [0.0332866057653704]},
    "adaptive_ece": {"n": 1, "mean": 0.0363536356148751, "std": null, "lo": null, "hi": null, "values": [0.0363536356148751]},
    "mce": {"n": 1, "mean": 0.3488495449225107, "std": null, "lo": null, "hi": null, "values": [0.3488495449225107]},
    "brier": {"n": 1, "mean": 0.4697349304327781, "std": null, "lo": null, "hi": null, "values": [0.4697349304327781]},
    "temperature": {"n": 1, "mean": 1.0646052662905536, "std": null, "lo": null, "hi": null, "values": [1.0646052662905536]},
    "ece_temp": {"n": 1, "mean": 0.0340101010641395, "std": null, "lo": null, "hi": null, "values": [0.0340101010641395]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.032904254828184, "std": null, "lo": null, "hi": null, "values": [0.032904254828184]},
    "mce_temp": {"n": 1, "mean": 0.3525947934321009, "std": null, "lo": null, "hi": null, "values": [0.3525947934321009]},
    "brier_temp": {"n": 1, "mean": 0.4690676959143746, "std": null, "lo": null, "hi": null, "values": [0.4690676959143746]}
   }},
  {"model": "kanres-deep", "task": "physionet2017-4class", "averaging": "macro", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.6411503071210004, "std": null, "lo": null, "hi": null, "values": [0.6411503071210004]},
    "auroc": {"n": 1, "mean": 0.9143775922438756, "std": null, "lo": null, "hi": null, "values": [0.9143775922438756]},
    "accuracy": {"n": 1, "mean": 0.779970215934475, "std": null, "lo": null, "hi": null, "values": [0.779970215934475]},
    "mcc": {"n": 1, "mean": 0.5917013689157004, "std": null, "lo": null, "hi": null, "values": [0.5917013689157004]},
    "precision": {"n": 1, "mean": 0.6567261908575627, "std": null, "lo": null, "hi": null, "values": [0.6567261908575627]},
    "recall": {"n": 1, "mean": 0.6310311755515212, "std": null, "lo": null, "hi": null, "values": [0.6310311755515212]},
    "ece": {"n": 1, "mean": 0.0229356514036167, "std": null, "lo": null, "hi": null, "values": [0.0229356514036167]},
    "adaptive_ece": {"n": 1, "mean": 0.0292556661882244, "std": null, "lo": null, "hi": null, "values": [0.0292556661882244]},
    "mce": {"n": 1, "mean": 0.3281933963298797, "std": null, "lo": null, "hi": null, "values": [0.3281933963298797]},
    "brier": {"n": 1, "mean": 0.314666367847759, "std": null, "lo": null, "hi": null, "values": [0.314666367847759]},
    "temperature": {"n": 1, "mean": 1.2079595622206425, "std": null, "lo": null, "hi": null, "values": [1.2079595622206425]},
    "ece_temp": {"n": 1, "mean": 0.023366113114901, "std": null, "lo": null, "hi": null, "values": [0.023366113114901]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0264715982748502, "std": null, "lo": null, "hi": null, "values": [0.0264715982748502]},
    "mce_temp": {"n": 1, "mean": 0.314680372439679, "std": null, "lo": null, "hi": null, "values": [0.314680372439679]},
    "brier_temp": {"n": 1, "mean": 0.3137751992535581, "std": null, "lo": null, "hi": null, "values": [0.3137751992535581]}
   }},
  {"model": "lstm", "task": "physionet2017-4class", "averaging": "macro", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.1850375234521576, "std": null, "lo": null, "hi": null, "values": [0.1850375234521576]},
    "auroc": {"n": 1, "mean": 0.51620952271031, "std": null, "lo": null, "hi": null, "values": [0.51620952271031]},
    "accuracy": {"n": 1, "mean": 0.5874906924795235, "std": null, "lo": null, "hi": null, "values": [0.5874906924795235]},
    "mcc": {"n": 1, "mean": 0.0, "std": null, "lo": null, "hi": null, "values": [0.0]},
    "precision": {"n": 1, "mean": 0.1468726731198808, "std": null, "lo": null, "hi": null, "values": [0.1468726731198808]},
    "recall": {"n": 1, "mean": 0.25, "std": null, "lo": null, "hi": null, "values": [0.25]},
    "ece": {"n": 1, "mean": 0.0015167095549321, "std": null, "lo": null, "hi": null, "values": [0.0015167095549321]},
    "adaptive_ece": {"n": 1, "mean": 0.0246306309394666, "std": null, "lo": null, "hi": null, "values": [0.0246306309394666]},
    "mce": {"n": 1, "mean": 0.1183419227600097, "std": null, "lo": null, "hi": null, "values": [0.1183419227600097]},
    "brier": {"n": 1, "mean": 0.5524747818379054, "std": null, "lo": null, "hi": null, "values": [0.5524747818379054]},
    "temperature": {"n": 1, "mean": 0.9853416542135434, "std": null, "lo": null, "hi": null, "values": [0.9853416542135434]},
    "ece_temp": {"n": 1, "mean": 0.003894941794668, "std": null, "lo": null, "hi": null, "values": [0.003894941794668]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0246738398470184, "std": null, "lo": null, "hi": null, "values": [0.0246738398470184]},
    "mce_temp": {"n": 1, "mean": 0.0039720748188559, "std": null, "lo": null, "hi": null, "values": [0.0039720748188559]},
    "brier_temp": {"n": 1, "mean": 0.552482430755466, "std": null, "lo": null, "hi": null, "values": [0.552482430755466]}
   }},
  {"model": "medformer", "task": "physionet2017-4class", "averaging": "macro", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.3111452254618632, "std": null, "lo": null, "hi": null, "values": [0.3111452254618632]},
    "auroc": {"n": 1, "mean": 0.7229192675842324, "std": null, "lo": null, "hi": null, "values": [0.7229192675842324]},
    "accuracy": {"n": 1, "mean": 0.62397617274758, "std": null, "lo": null, "hi": null, "values": [0.62397617274758]},
    "mcc": {"n": 1, "mean": 0.2144385974886049, "std": null, "lo": null, "hi": null, "values": [0.2144385974886049]},
    "precision": {"n": 1, "mean": 0.3817540484207151, "std": null, "lo": null, "hi": null, "values": [0.3817540484207151]},
    "recall": {"n": 1, "mean": 0.316727802875159, "std": null, "lo": null, "hi": null, "values": [0.316727802875159]},
    "ece": {"n": 1, "mean": 0.0460446173476356, "std": null, "lo": null, "hi": null, "values": [0.0460446173476356]},
    "adaptive_ece": {"n": 1, "mean": 0.0479448210345393, "std": null, "lo": null, "hi": null, "values": [0.0479448210345393]},
    "mce": {"n": 1, "mean": 0.3209609389305115, "std": null, "lo": null, "hi": null, "values": [0.3209609389305115]},
    "brier": {"n": 1, "mean": 0.5102098241727479, "std": null, "lo": null, "hi": null, "values": [0.5102098241727479]},
    "temperature": {"n": 1, "mean": 0.9045213850548957, "std": null, "lo": null, "hi": null, "values": [0.9045213850548957]},
    "ece_temp": {"n": 1, "mean": 0.0318566049222582, "std": null, "lo": null, "hi": null, "values": [0.0318566049222582]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0364511433784014, "std": null, "lo": null, "hi": null, "values": [0.0364511433784014]},
    "mce_temp": {"n": 1, "mean": 0.3253442630001976, "std": null, "lo": null, "hi": null, "values": [0.3253442630001976]},
    "brier_temp": {"n": 1, "mean": 0.5081874287685741, "std": null, "lo": null, "hi": null, "values": [0.5081874287685741]}
   }},
  {"model": "mlp-wang", "task": "physionet2017-4class", "averaging": "macro", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.203066342627829, "std": null, "lo": null, "hi": null, "values": [0.203066342627829]},
    "auroc": {"n": 1, "mean": 0.5504240629539944, "std": null, "lo": null, "hi": null, "values": [0.5504240629539944]},
    "accuracy": {"n": 1, "mean": 0.5889798957557707, "std": null, "lo": null, "hi": null, "values": [0.5889798957557707]},
    "mcc": {"n": 1, "mean": 0.0526192962151163, "std": null, "lo": null, "hi": null, "values": [0.0526192962151163]},
    "precision": {"n": 1, "mean": 0.2642851372094902, "std": null, "lo": null, "hi": null, "values": [0.2642851372094902]},
    "recall": {"n": 1, "mean": 0.2553693447952829, "std": null, "lo": null, "hi": null, "values": [0.2553693447952829]},
    "ece": {"n": 1, "mean": 0.0413822059700583, "std": null, "lo": null, "hi": null, "values": [0.0413822059700583]},
    "adaptive_ece": {"n": 1, "mean": 0.0473949516653925, "std": null, "lo": null, "hi": null, "values": [0.0473949516653925]},
    "mce": {"n": 1, "mean": 0.3382347921530405, "std": null, "lo": null, "hi": null, "values": [0.3382347921530405]},
    "brier": {"n": 1, "mean": 0.5521018605019342, "std": null, "lo": null, "hi": null, "values": [0.5521018605019342]},
    "temperature": {"n": 1, "mean": 1.0880025739950254, "std": null, "lo": null, "hi": null, "values": [1.0880025739950254]},
    "ece_temp": {"n": 1, "mean": 0.0548168618466669, "std": null, "lo": null, "hi": null, "values": [0.0548168618466669]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0512835057231469, "std": null, "lo": null, "hi": null, "values": [0.0512835057231469]},
    "mce_temp": {"n": 1, "mean": 0.3501274074819291, "std": null, "lo": null, "hi": null, "values": [0.3501274074819291]},
    "brier_temp": {"n": 1, "mean": 0.5523573564461393, "std": null, "lo": null, "hi": null, "values": [0.5523573564461393]}
   }},
  {"model": "resnet", "task": "physionet2017-4class", "averaging": "macro", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.6634654112065299, "std": null, "lo": null, "hi": null, "values": [0.6634654112065299]},
    "auroc": {"n": 1, "mean": 0.91389644070265, "std": null, "lo": null, "hi": null, "values": [0.91389644070265]},
    "accuracy": {"n": 1, "mean": 0.7892777364110201, "std": null, "lo": null, "hi": null, "values": [0.7892777364110201]},
    "mcc": {"n": 1, "mean": 0.6077319184349689, "std": null, "lo": null, "hi": null, "values": [0.6077319184349689]},
    "precision": {"n": 1, "mean": 0.6920924710397155, "std": null, "lo": null, "hi": null, "values": [0.6920924710397155]},
    "recall": {"n": 1, "mean": 0.6412506629617486, "std": null, "lo": null, "hi": null, "values": [0.6412506629617486]},
    "ece": {"n": 1, "mean": 0.0199121561231492, "std": null, "lo": null, "hi": null, "values": [0.0199121561231492]},
    "adaptive_ece": {"n": 1, "mean": 0.0268774988796847, "std": null, "lo": null, "hi": null, "values": [0.0268774988796847]},
    "mce": {"n": 1, "mean": 0.0756727793277838, "std": null, "lo": null, "hi": null, "values": [0.0756727793277838]},
    "brier": {"n": 1, "mean": 0.3050765048647064, "std": null, "lo": null, "hi": null, "values": [0.3050765048647064]},
    "temperature": {"n": 1, "mean": 1.12114639807714, "std": null, "lo": null, "hi": null, "values": [1.12114639807714]},
    "ece_temp": {"n": 1, "mean": 0.0365603255892785, "std": null, "lo": null, "hi": null, "values": [0.0365603255892785]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0340631143088602, "std": null, "lo": null, "hi": null, "values": [0.0340631143088602]},
    "mce_temp": {"n": 1, "mean": 0.3241918754230186, "std": null, "lo": null, "hi": null, "values": [0.3241918754230186]},
    "brier_temp": {"n": 1, "mean": 0.306059796477478, "std": null, "lo": null, "hi": null, "values": [0.306059796477478]}
   }},
  {"model": "resnet-hannun", "task": "physionet2017-4class", "averaging": "macro", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.679439050747923, "std": null, "lo": null, "hi": null, "values": [0.679439050747923]},
    "auroc": {"n": 1, "mean": 0.9153756754761004, "std": null, "lo": null, "hi": null, "values": [0.9153756754761004]},
    "accuracy": {"n": 1, "mean": 0.8034251675353685, "std": null, "lo": null, "hi": null, "values": [0.8034251675353685]},
    "mcc": {"n": 1, "mean": 0.6418233160683225, "std": null, "lo": null, "hi": null, "values": [0.6418233160683225]},
    "precision": {"n": 1, "mean": 0.6721288265312344, "std": null, "lo": null, "hi": null, "values": [0.6721288265312344]},
    "recall": {"n": 1, "mean": 0.6945986619648736, "std": null, "lo": null, "hi": null, "values": [0.6945986619648736]},
    "ece": {"n": 1, "mean": 0.0424603442491565, "std": null, "lo": null, "hi": null, "values": [0.0424603442491565]},
    "adaptive_ece": {"n": 1, "mean": 0.0426684074653056, "std": null, "lo": null, "hi": null, "values": [0.0426684074653056]},
    "mce": {"n": 1, "mean": 0.7047852277755737, "std": null, "lo": null, "hi": null, "values": [0.7047852277755737]},
    "brier": {"n": 1, "mean": 0.297821360795595, "std": null, "lo": null, "hi": null, "values": [0.297821360795595]},
    "temperature": {"n": 1, "mean": 0.8787373533844887, "std": null, "lo": null, "hi": null, "values": [0.8787373533844887]},
    "ece_temp": {"n": 1, "mean": 0.0236859401538858, "std": null, "lo": null, "hi": null, "values": [0.0236859401538858]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0219717655021822, "std": null, "lo": null, "hi": null, "values": [0.0219717655021822]},
    "mce_temp": {"n": 1, "mean": 0.698620348351845, "std": null, "lo": null, "hi": null, "values": [0.698620348351845]},
    "brier_temp": {"n": 1, "mean": 0.2956371563610828, "std": null, "lo": null, "hi": null, "values": [0.2956371563610828]}
   }},
  {"model": "resnet-wang", "task": "physionet2017-4class", "averaging": "macro", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.4282884641600417, "std": null, "lo": null, "hi": null, "values": [0.4282884641600417]},
    "auroc": {"n": 1, "mean": 0.7926294460420362, "std": null, "lo": null, "hi": null, "values": [0.7926294460420362]},
    "accuracy": {"n": 1, "mean": 0.6470588235294118, "std": null, "lo": null, "hi": null, "values": [0.6470588235294118]},
    "mcc": {"n": 1, "mean": 0.2902655265109951, "std": null, "lo": null, "hi": null, "values": [0.2902655265109951]},
    "precision": {"n": 1, "mean": 0.607681779649659, "std": null, "lo": null, "hi": null, "values": [0.607681779649659]},
    "recall": {"n": 1, "mean": 0.3942633029155594, "std": null, "lo": null, "hi": null, "values": [0.3942633029155594]},
    "ece": {"n": 1, "mean": 0.0441054824853422, "std": null, "lo": null, "hi": null, "values": [0.0441054824853422]},
    "adaptive_ece": {"n": 1, "mean": 0.0501439221319829, "std": null, "lo": null, "hi": null, "values": [0.0501439221319829]},
    "mce": {"n": 1, "mean": 0.3189395070075989, "std": null, "lo": null, "hi": null, "values": [0.3189395070075989]},
    "brier": {"n": 1, "mean": 0.4700314885441051, "std": null, "lo": null, "hi": null, "values": [0.4700314885441051]},
    "temperature": {"n": 1, "mean": 1.14628712571144, "std": null, "lo": null, "hi": null, "values": [1.14628712571144]},
    "ece_temp": {"n": 1, "mean": 0.0315230676979048, "std": null, "lo": null, "hi": null, "values": [0.0315230676979048]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0362957731982102, "std": null, "lo": null, "hi": null, "values": [0.0362957731982102]},
    "mce_temp": {"n": 1, "mean": 0.3128634818092198, "std": null, "lo": null, "hi": null, "values": [0.3128634818092198]},
    "brier_temp": {"n": 1, "mean": 0.467843241235333, "std": null, "lo": null, "hi": null, "values": [0.467843241235333]}
   }},
  {"model": "rocket", "task": "physionet2017-4class", "averaging": "macro", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.555428040447167, "std": null, "lo": null, "hi": null, "values": [0.555428040447167]},
    "auroc": {"n": 1, "mean": 0.8516748614043518, "std": null, "lo": null, "hi": null, "values": [0.8516748614043518]},
    "accuracy": {"n": 1, "mean": 0.7170513775130305, "std": null, "lo": null, "hi": null, "values": [0.7170513775130305]},
    "mcc": {"n": 1, "mean": 0.4660356210229475, "std": null, "lo": null, "hi": null, "values": [0.4660356210229475]},
    "precision": {"n": 1, "mean": 0.5896020424915983, "std": null, "lo": null, "hi": null, "values": [0.5896020424915983]},
    "recall": {"n": 1, "mean": 0.5358886331951538, "std": null, "lo": null, "hi": null, "values": [0.5358886331951538]},
    "ece": {"n": 1, "mean": 0.0332421558900103, "std": null, "lo": null, "hi": null, "values": [0.0332421558900103]},
    "adaptive_ece": {"n": 1, "mean": 0.0337793127046266, "std": null, "lo": null, "hi": null, "values": [0.0337793127046266]},
    "mce": {"n": 1, "mean": 0.1231421689902033, "std": null, "lo": null, "hi": null, "values": [0.1231421689902033]},
    "brier": {"n": 1, "mean": 0.3993740265558728, "std": null, "lo": null, "hi": null, "values": [0.3993740265558728]},
    "temperature": {"n": 1, "mean": 1.2244066580483004, "std": null, "lo": null, "hi": null, "values": [1.2244066580483004]},
    "ece_temp": {"n": 1, "mean": 0.0263496356704698, "std": null, "lo": null, "hi": null, "values": [0.0263496356704698]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0236856867308806, "std": null, "lo": null, "hi": null, "values": [0.0236856867308806]},
    "mce_temp": {"n": 1, "mean": 0.1027043668976653, "std": null, "lo": null, "hi": null, "values": [0.1027043668976653]},
    "brier_temp": {"n": 1, "mean": 0.3978149637047184, "std": null, "lo": null, "hi": null, "values": [0.3978149637047184]}
   }},
  {"model": "se-resnet", "task": "physionet2017-4class", "averaging": "macro", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.683725204428136, "std": null, "lo": null, "hi": null, "values": [0.683725204428136]},
    "auroc": {"n": 1, "mean": 0.9267677309391996, "std": null, "lo": null, "hi": null, "values": [0.9267677309391996]},
    "accuracy": {"n": 1, "mean": 0.8071481757259866, "std": null, "lo": null, "hi": null, "values": [0.8071481757259866]},
    "mcc": {"n": 1, "mean": 0.6438526272340734, "std": null, "lo": null, "hi": null, "values": [0.6438526272340734]},
    "precision": {"n": 1, "mean": 0.6943098632745186, "std": null, "lo": null, "hi": null, "values": [0.6943098632745186]},
    "recall": {"n": 1, "mean": 0.6775613096699226, "std": null, "lo": null, "hi": null, "values": [0.6775613096699226]},
    "ece": {"n": 1, "mean": 0.0321374523910407, "std": null, "lo": null, "hi": null, "values": [0.0321374523910407]},
    "adaptive_ece": {"n": 1, "mean": 0.0322321870281726, "std": null, "lo": null, "hi": null, "values": [0.0322321870281726]},
    "mce": {"n": 1, "mean": 0.1425067216492771, "std": null, "lo": null, "hi": null, "values": [0.1425067216492771]},
    "brier": {"n": 1, "mean": 0.2854619804187129, "std": null, "lo": null, "hi": null, "values": [0.2854619804187129]},
    "temperature": {"n": 1, "mean": 1.1903711508895936, "std": null, "lo": null, "hi": null, "values": [1.1903711508895936]},
    "ece_temp": {"n": 1, "mean": 0.0244411941901469, "std": null, "lo": null, "hi": null, "values": [0.0244411941901469]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0234534952251791, "std": null, "lo": null, "hi": null, "values": [0.0234534952251791]},
    "mce_temp": {"n": 1, "mean": 0.0969526311879632, "std": null, "lo": null, "hi": null, "values": [0.0969526311879632]},
    "brier_temp": {"n": 1, "mean": 0.2846545436916064, "std": null, "lo": null, "hi": null, "values": [0.2846545436916064]}
   }},
  {"model": "simple-cnn", "task": "physionet2017-4class", "averaging": "macro", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.4902715942671874, "std": null, "lo": null, "hi": null, "values": [0.4902715942671874]},
    "auroc": {"n": 1, "mean": 0.8294962307324303, "std": null, "lo": null, "hi": null, "values": [0.8294962307324303]},
    "accuracy": {"n": 1, "mean": 0.6708860759493671, "std": null, "lo": null, "hi": null, "values": [0.6708860759493671]},
    "mcc": {"n": 1, "mean": 0.3685700892104239, "std": null, "lo": null, "hi": null, "values": [0.3685700892104239]},
    "precision": {"n": 1, "mean": 0.5638270709508201, "std": null, "lo": null, "hi": null, "values": [0.5638270709508201]},
    "recall": {"n": 1, "mean": 0.4785212981328632, "std": null, "lo": null, "hi": null, "values": [0.4785212981328632]},
    "ece": {"n": 1, "mean": 0.0214777194845614, "std": null, "lo": null, "hi": null, "values": [0.0214777194845614]},
    "adaptive_ece": {"n": 1, "mean": 0.0221577115073988, "std": null, "lo": null, "hi": null, "values": [0.0221577115073988]},
    "mce": {"n": 1, "mean": 0.3167593777179718, "std": null, "lo": null, "hi": null, "values": [0.3167593777179718]},
    "brier": {"n": 1, "mean": 0.4441203692223364, "std": null, "lo": null, "hi": null, "values": [0.4441203692223364]},
    "temperature": {"n": 1, "mean": 1.0505389722696332, "std": null, "lo": null, "hi": null, "values": [1.0505389722696332]},
    "ece_temp": {"n": 1, "mean": 0.025534504225627, "std": null, "lo": null, "hi": null, "values": [0.025534504225627]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0210416036712498, "std": null, "lo": null, "hi": null, "values": [0.0210416036712498]},
    "mce_temp": {"n": 1, "mean": 0.3140932216625229, "std": null, "lo": null, "hi": null, "values": [0.3140932216625229]},
    "brier_temp": {"n": 1, "mean": 0.443881237899457, "std": null, "lo": null, "hi": null, "values": [0.443881237899457]}
   }},
  {"model": "spectrogram-cnn", "task": "physionet2017-4class", "averaging": "macro", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.6841141471730359, "std": null, "lo": null, "hi": null, "values": [0.6841141471730359]},
    "auroc": {"n": 1, "mean": 0.9178592098033276, "std": null, "lo": null, "hi": null, "values": [0.9178592098033276]},
    "accuracy": {"n": 1, "mean": 0.8060312732688012, "std": null, "lo": null, "hi": null, "values": [0.8060312732688012]},
    "mcc": {"n": 1, "mean": 0.6412805010170892, "std": null, "lo": null, "hi": null, "values": [0.6412805010170892]},
    "precision": {"n": 1, "mean": 0.6977991475211941, "std": null, "lo": null, "hi": null, "values": [0.6977991475211941]},
    "recall": {"n": 1, "mean": 0.6750320763070804, "std": null, "lo": null, "hi": null, "values": [0.6750320763070804]},
    "ece": {"n": 1, "mean": 0.0363869788326262, "std": null, "lo": null, "hi": null, "values": [0.0363869788326262]},
    "adaptive_ece": {"n": 1, "mean": 0.0364409269167253, "std": null, "lo": null, "hi": null, "values": [0.0364409269167253]},
    "mce": {"n": 1, "mean": 0.3034154325723648, "std": null, "lo": null, "hi": null, "values": [0.3034154325723648]},
    "brier": {"n": 1, "mean": 0.2955308089105865, "std": null, "lo": null, "hi": null, "values": [0.2955308089105865]},
    "temperature": {"n": 1, "mean": 0.9952833636731652, "std": null, "lo": null, "hi": null, "values": [0.9952833636731652]},
    "ece_temp": {"n": 1, "mean": 0.0356099949984025, "std": null, "lo": null, "hi": null, "values": [0.0356099949984025]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0351349248737368, "std": null, "lo": null, "hi": null, "values": [0.0351349248737368]},
    "mce_temp": {"n": 1, "mean": 0.3036804821601555, "std": null, "lo": null, "hi": null, "values": [0.3036804821601555]},
    "brier_temp": {"n": 1, "mean": 0.2954608007501675, "std": null, "lo": null, "hi": null, "values": [0.2954608007501675]}
   }},
  {"model": "tcn", "task": "physionet2017-4class", "averaging": "macro", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.6194032864180222, "std": null, "lo": null, "hi": null, "values": [0.6194032864180222]},
    "auroc": {"n": 1, "mean": 0.9032192783307484, "std": null, "lo": null, "hi": null, "values": [0.9032192783307484]},
    "accuracy": {"n": 1, "mean": 0.7732688011913627, "std": null, "lo": null, "hi": null, "values": [0.7732688011913627]},
    "mcc": {"n": 1, "mean": 0.5834468090487869, "std": null, "lo": null, "hi": null, "values": [0.5834468090487869]},
    "precision": {"n": 1, "mean": 0.6409719317677189, "std": null, "lo": null, "hi": null, "values": [0.6409719317677189]},
    "recall": {"n": 1, "mean": 0.6188395228392984, "std": null, "lo": null, "hi": null, "values": [0.6188395228392984]},
    "ece": {"n": 1, "mean": 0.0252204091244461, "std": null, "lo": null, "hi": null, "values": [0.0252204091244461]},
    "adaptive_ece": {"n": 1, "mean": 0.0283732586113801, "std": null, "lo": null, "hi": null, "values": [0.0283732586113801]},
    "mce": {"n": 1, "mean": 0.1861160825937986, "std": null, "lo": null, "hi": null, "values": [0.1861160825937986]},
    "brier": {"n": 1, "mean": 0.3287552802194826, "std": null, "lo": null, "hi": null, "values": [0.3287552802194826]},
    "temperature": {"n": 1, "mean": 1.1215608120582974, "std": null, "lo": null, "hi": null, "values": [1.1215608120582974]},
    "ece_temp": {"n": 1, "mean": 0.0335195365937262, "std": null, "lo": null, "hi": null, "values": [0.0335195365937262]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0370023879654661, "std": null, "lo": null, "hi": null, "values": [0.0370023879654661]},
    "mce_temp": {"n": 1, "mean": 0.1233508331519214, "std": null, "lo": null, "hi": null, "values": [0.1233508331519214]},
    "brier_temp": {"n": 1, "mean": 0.3297983726605257, "std": null, "lo": null, "hi": null, "values": [0.3297983726605257]}
   }},
  {"model": "transformer", "task": "physionet2017-4class", "averaging": "macro", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.1850375234521576, "std": null, "lo": null, "hi": null, "values": [0.1850375234521576]},
    "auroc": {"n": 1, "mean": 0.4927558651815606, "std": null, "lo": null, "hi": null, "values": [0.4927558651815606]},
    "accuracy": {"n": 1, "mean": 0.5874906924795235, "std": null, "lo": null, "hi": null, "values": [0.5874906924795235]},
    "mcc": {"n": 1, "mean": 0.0, "std": null, "lo": null, "hi": null, "values": [0.0]},
    "precision": {"n": 1, "mean": 0.1468726731198808, "std": null, "lo": null, "hi": null, "values": [0.1468726731198808]},
    "recall": {"n": 1, "mean": 0.25, "std": null, "lo": null, "hi": null, "values": [0.25]},
    "ece": {"n": 1, "mean": 0.0013180285004298, "std": null, "lo": null, "hi": null, "values": [0.0013180285004298]},
    "adaptive_ece": {"n": 1, "mean": 0.0453504872073545, "std": null, "lo": null, "hi": null, "values": [0.0453504872073545]},
    "mce": {"n": 1, "mean": 0.0013180285004298, "std": null, "lo": null, "hi": null, "values": [0.0013180285004298]},
    "brier": {"n": 1, "mean": 0.5522486481632896, "std": null, "lo": null, "hi": null, "values": [0.5522486481632896]},
    "temperature": {"n": 1, "mean": 0.9936484953882164, "std": null, "lo": null, "hi": null, "values": [0.9936484953882164]},
    "ece_temp": {"n": 1, "mean": 0.0003604216759657, "std": null, "lo": null, "hi": null, "values": [0.0003604216759657]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0487385439757526, "std": null, "lo": null, "hi": null, "values": [0.0487385439757526]},
    "mce_temp": {"n": 1, "mean": 0.0003604216759657, "std": null, "lo": null, "hi": null, "values": [0.0003604216759657]},
    "brier_temp": {"n": 1, "mean": 0.5522440875308413, "std": null, "lo": null, "hi": null, "values": [0.5522440875308413]}
   }},
  {"model": "xresnet", "task": "physionet2017-4class", "averaging": "macro", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.6234927875562412, "std": null, "lo": null, "hi": null, "values": [0.6234927875562412]},
    "auroc": {"n": 1, "mean": 0.9077853703414728, "std": null, "lo": null, "hi": null, "values": [0.9077853703414728]},
    "accuracy": {"n": 1, "mean": 0.7673119880863738, "std": null, "lo": null, "hi": null, "values": [0.7673119880863738]},
    "mcc": {"n": 1, "mean": 0.5623839645233736, "std": null, "lo": null, "hi": null, "values": [0.5623839645233736]},
    "precision": {"n": 1, "mean": 0.7491234703528398, "std": null, "lo": null, "hi": null, "values": [0.7491234703528398]},
    "recall": {"n": 1, "mean": 0.5794962063235449, "std": null, "lo": null, "hi": null, "values": [0.5794962063235449]},
    "ece": {"n": 1, "mean": 0.0236853396857042, "std": null, "lo": null, "hi": null, "values": [0.0236853396857042]},
    "adaptive_ece": {"n": 1, "mean": 0.0328910602665799, "std": null, "lo": null, "hi": null, "values": [0.0328910602665799]},
    "mce": {"n": 1, "mean": 0.1333905934717096, "std": null, "lo": null, "hi": null, "values": [0.1333905934717096]},
    "brier": {"n": 1, "mean": 0.326739241028877, "std": null, "lo": null, "hi": null, "values": [0.326739241028877]},
    "temperature": {"n": 1, "mean": 1.1344912512176215, "std": null, "lo": null, "hi": null, "values": [1.1344912512176215]},
    "ece_temp": {"n": 1, "mean": 0.0320786117203278, "std": null, "lo": null, "hi": null, "values": [0.0320786117203278]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0283796549391308, "std": null, "lo": null, "hi": null, "values": [0.0283796549391308]},
    "mce_temp": {"n": 1, "mean": 0.1104349432301984, "std": null, "lo": null, "hi": null, "values": [0.1104349432301984]},
    "brier_temp": {"n": 1, "mean": 0.3260902600817782, "std": null, "lo": null, "hi": null, "values": [0.3260902600817782]}
   }},
  {"model": "xresnet1d-benchmark", "task": "physionet2017-4class", "averaging": "macro", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.652019106672572, "std": null, "lo": null, "hi": null, "values": [0.652019106672572]},
    "auroc": {"n": 1, "mean": 0.908951963272596, "std": null, "lo": null, "hi": null, "values": [0.908951963272596]},
    "accuracy": {"n": 1, "mean": 0.7885331347728965, "std": null, "lo": null, "hi": null, "values": [0.7885331347728965]},
    "mcc": {"n": 1, "mean": 0.6112364848962436, "std": null, "lo": null, "hi": null, "values": [0.6112364848962436]},
    "precision": {"n": 1, "mean": 0.6723167755991589, "std": null, "lo": null, "hi": null, "values": [0.6723167755991589]},
    "recall": {"n": 1, "mean": 0.6472988502574266, "std": null, "lo": null, "hi": null, "values": [0.6472988502574266]},
    "ece": {"n": 1, "mean": 0.0343291858018376, "std": null, "lo": null, "hi": null, "values": [0.0343291858018376]},
    "adaptive_ece": {"n": 1, "mean": 0.0324141806572653, "std": null, "lo": null, "hi": null, "values": [0.0324141806572653]},
    "mce": {"n": 1, "mean": 0.332518070936203, "std": null, "lo": null, "hi": null, "values": [0.332518070936203]},
    "brier": {"n": 1, "mean": 0.3169275979628581, "std": null, "lo": null, "hi": null, "values": [0.3169275979628581]},
    "temperature": {"n": 1, "mean": 1.204852098826143, "std": null, "lo": null, "hi": null, "values": [1.204852098826143]},
    "ece_temp": {"n": 1, "mean": 0.0181401650164985, "std": null, "lo": null, "hi": null, "values": [0.0181401650164985]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0243593143391799, "std": null, "lo": null, "hi": null, "values": [0.0243593143391799]},
    "mce_temp": {"n": 1, "mean": 0.3231047847483771, "std": null, "lo": null, "hi": null, "values": [0.3231047847483771]},
    "brier_temp": {"n": 1, "mean": 0.3150751914296749, "std": null, "lo": null, "hi": null, "values": [0.3150751914296749]}
   }},
  {"model": "xresnet1d101", "task": "physionet2017-4class", "averaging": "macro", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.6540794846481978, "std": null, "lo": null, "hi": null, "values": [0.6540794846481978]},
    "auroc": {"n": 1, "mean": 0.9035447326287572, "std": null, "lo": null, "hi": null, "values": [0.9035447326287572]},
    "accuracy": {"n": 1, "mean": 0.7862993298585257, "std": null, "lo": null, "hi": null, "values": [0.7862993298585257]},
    "mcc": {"n": 1, "mean": 0.6037942738312969, "std": null, "lo": null, "hi": null, "values": [0.6037942738312969]},
    "precision": {"n": 1, "mean": 0.6746316956125578, "std": null, "lo": null, "hi": null, "values": [0.6746316956125578]},
    "recall": {"n": 1, "mean": 0.6378617071249265, "std": null, "lo": null, "hi": null, "values": [0.6378617071249265]},
    "ece": {"n": 1, "mean": 0.0431727592348608, "std": null, "lo": null, "hi": null, "values": [0.0431727592348608]},
    "adaptive_ece": {"n": 1, "mean": 0.0426703157792407, "std": null, "lo": null, "hi": null, "values": [0.0426703157792407]},
    "mce": {"n": 1, "mean": 0.3270940780639648, "std": null, "lo": null, "hi": null, "values": [0.3270940780639648]},
    "brier": {"n": 1, "mean": 0.3118640835831467, "std": null, "lo": null, "hi": null, "values": [0.3118640835831467]},
    "temperature": {"n": 1, "mean": 1.3396258738021876, "std": null, "lo": null, "hi": null, "values": [1.3396258738021876]},
    "ece_temp": {"n": 1, "mean": 0.0274805922251955, "std": null, "lo": null, "hi": null, "values": [0.0274805922251955]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0274736464054867, "std": null, "lo": null, "hi": null, "values": [0.0274736464054867]},
    "mce_temp": {"n": 1, "mean": 0.3127508604944107, "std": null, "lo": null, "hi": null, "values": [0.3127508604944107]},
    "brier_temp": {"n": 1, "mean": 0.3092530018295408, "std": null, "lo": null, "hi": null, "values": [0.3092530018295408]}
   }},
  {"model": "convnext-v2", "task": "mitbih-afdb-binary", "averaging": "binary", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.0, "std": null, "lo": null, "hi": null, "values": [0.0]},
    "auroc": {"n": 1, "mean": 0.5, "std": null, "lo": null, "hi": null, "values": [0.5]},
    "accuracy": {"n": 1, "mean": 0.6478815860945139, "std": null, "lo": null, "hi": null, "values": [0.6478815860945139]},
    "mcc": {"n": 1, "mean": 0.0, "std": null, "lo": null, "hi": null, "values": [0.0]},
    "precision": {"n": 1, "mean": 0.0, "std": null, "lo": null, "hi": null, "values": [0.0]},
    "recall": {"n": 1, "mean": 0.0, "std": null, "lo": null, "hi": null, "values": [0.0]},
    "ece": {"n": 1, "mean": 0.0660085268217477, "std": null, "lo": null, "hi": null, "values": [0.0660085268217477]},
    "adaptive_ece": {"n": 1, "mean": 0.4538185451420541, "std": null, "lo": null, "hi": null, "values": [0.4538185451420541]},
    "mce": {"n": 1, "mean": 0.0660085268217477, "std": null, "lo": null, "hi": null, "values": [0.0660085268217477]},
    "brier": {"n": 1, "mean": 0.4649763242146965, "std": null, "lo": null, "hi": null, "values": [0.4649763242146965]},
    "temperature": {"n": 1, "mean": 1.0578260248963027, "std": null, "lo": null, "hi": null, "values": [1.0578260248963027]},
    "ece_temp": {"n": 1, "mean": 0.0704100463364006, "std": null, "lo": null, "hi": null, "values": [0.0704100463364006]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.4241997487131798, "std": null, "lo": null, "hi": null, "values": [0.4241997487131798]},
    "mce_temp": {"n": 1, "mean": 0.0704100463364006, "std": null, "lo": null, "hi": null, "values": [0.0704100463364006]},
    "brier_temp": {"n": 1, "mean": 0.4661772224041584, "std": null, "lo": null, "hi": null, "values": [0.4661772224041584]}
   }},
  {"model": "crnn", "task": "mitbih-afdb-binary", "averaging": "binary", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.9771260997067448, "std": null, "lo": null, "hi": null, "values": [0.9771260997067448]},
    "auroc": {"n": 1, "mean": 0.998349473369728, "std": null, "lo": null, "hi": null, "values": [0.998349473369728]},
    "accuracy": {"n": 1, "mean": 0.984111895708854, "std": null, "lo": null, "hi": null, "values": [0.984111895708854]},
    "mcc": {"n": 1, "mean": 0.965176912615629, "std": null, "lo": null, "hi": null, "values": [0.965176912615629]},
    "precision": {"n": 1, "mean": 0.9908802537668516, "std": null, "lo": null, "hi": null, "values": [0.9908802537668516]},
    "recall": {"n": 1, "mean": 0.9637485537986888, "std": null, "lo": null, "hi": null, "values": [0.9637485537986888]},
    "ece": {"n": 1, "mean": 0.0094095248767437, "std": null, "lo": null, "hi": null, "values": [0.0094095248767437]},
    "adaptive_ece": {"n": 1, "mean": 0.0090443098159797, "std": null, "lo": null, "hi": null, "values": [0.0090443098159797]},
    "mce": {"n": 1, "mean": 0.1475214759508768, "std": null, "lo": null, "hi": null, "values": [0.1475214759508768]},
    "brier": {"n": 1, "mean": 0.028540328627472, "std": null, "lo": null, "hi": null, "values": [0.028540328627472]},
    "temperature": {"n": 1, "mean": 2.17772149448971, "std": null, "lo": null, "hi": null, "values": [2.17772149448971]},
    "ece_temp": {"n": 1, "mean": 0.0346240527355532, "std": null, "lo": null, "hi": null, "values": [0.0346240527355532]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0346240527355531, "std": null, "lo": null, "hi": null, "values": [0.0346240527355531]},
    "mce_temp": {"n": 1, "mean": 0.3373007767657461, "std": null, "lo": null, "hi": null, "values": [0.3373007767657461]},
    "brier_temp": {"n": 1, "mean": 0.0302631349105183, "std": null, "lo": null, "hi": null, "values": [0.0302631349105183]}
   }},
  {"model": "deep-res-cnn", "task": "mitbih-afdb-binary", "averaging": "binary", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.9759178839320964, "std": null, "lo": null, "hi": null, "values": [0.9759178839320964]},
    "auroc": {"n": 1, "mean": 0.9712560290215916, "std": null, "lo": null, "hi": null, "values": [0.9712560290215916]},
    "accuracy": {"n": 1, "mean": 0.983432916892993, "std": null, "lo": null, "hi": null, "values": [0.983432916892993]},
    "mcc": {"n": 1, "mean": 0.9639300793162088, "std": null, "lo": null, "hi": null, "values": [0.9639300793162088]},
    "precision": {"n": 1, "mean": 0.9995956328346138, "std": null, "lo": null, "hi": null, "values": [0.9995956328346138]},
    "recall": {"n": 1, "mean": 0.9533359043578866, "std": null, "lo": null, "hi": null, "values": [0.9533359043578866]},
    "ece": {"n": 1, "mean": 0.0076202123159303, "std": null, "lo": null, "hi": null, "values": [0.0076202123159303]},
    "adaptive_ece": {"n": 1, "mean": 0.0219404200362097, "std": null, "lo": null, "hi": null, "values": [0.0219404200362097]},
    "mce": {"n": 1, "mean": 0.2294831037521362, "std": null, "lo": null, "hi": null, "values": [0.2294831037521362]},
    "brier": {"n": 1, "mean": 0.0322964111187439, "std": null, "lo": null, "hi": null, "values": [0.0322964111187439]},
    "temperature": {"n": 1, "mean": 1.7376774898523186, "std": null, "lo": null, "hi": null, "values": [1.7376774898523186]},
    "ece_temp": {"n": 1, "mean": 0.0411952910554552, "std": null, "lo": null, "hi": null, "values": [0.0411952910554552]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.060667572322779, "std": null, "lo": null, "hi": null, "values": [0.060667572322779]},
    "mce_temp": {"n": 1, "mean": 0.2947986317422043, "std": null, "lo": null, "hi": null, "values": [0.2947986317422043]},
    "brier_temp": {"n": 1, "mean": 0.0415681461555002, "std": null, "lo": null, "hi": null, "values": [0.0415681461555002]}
   }},
  {"model": "dualnet", "task": "mitbih-afdb-binary", "averaging": "binary", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.966789667896679, "std": null, "lo": null, "hi": null, "values": [0.966789667896679]},
    "auroc": {"n": 1, "mean": 0.9954345587894724, "std": null, "lo": null, "hi": null, "values": [0.9954345587894724]},
    "accuracy": {"n": 1, "mean": 0.9767789244975555, "std": null, "lo": null, "hi": null, "values": [0.9767789244975555]},
    "mcc": {"n": 1, "mean": 0.9489974549445584, "std": null, "lo": null, "hi": null, "values": [0.9489974549445584]},
    "precision": {"n": 1, "mean": 0.9737871674491392, "std": null, "lo": null, "hi": null, "values": [0.9737871674491392]},
    "recall": {"n": 1, "mean": 0.959892016968762, "std": null, "lo": null, "hi": null, "values": [0.959892016968762]},
    "ece": {"n": 1, "mean": 0.0127419829206709, "std": null, "lo": null, "hi": null, "values": [0.0127419829206709]},
    "adaptive_ece": {"n": 1, "mean": 0.011450695269526, "std": null, "lo": null, "hi": null, "values": [0.011450695269526]},
    "mce": {"n": 1, "mean": 0.0805968975319582, "std": null, "lo": null, "hi": null, "values": [0.0805968975319582]},
    "brier": {"n": 1, "mean": 0.0404732487591456, "std": null, "lo": null, "hi": null, "values": [0.0404732487591456]},
    "temperature": {"n": 1, "mean": 2.4044312673999295, "std": null, "lo": null, "hi": null, "values": [2.4044312673999295]},
    "ece_temp": {"n": 1, "mean": 0.0362319502178495, "std": null, "lo": null, "hi": null, "values": [0.0362319502178495]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0362319502178495, "std": null, "lo": null, "hi": null, "values": [0.0362319502178495]},
    "mce_temp": {"n": 1, "mean": 0.1373975928278602, "std": null, "lo": null, "hi": null, "values": [0.1373975928278602]},
    "brier_temp": {"n": 1, "mean": 0.0435992472471757, "std": null, "lo": null, "hi": null, "values": [0.0435992472471757]}
   }},
  {"model": "fcn-wang", "task": "mitbih-afdb-binary", "averaging": "binary", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.707274490785645, "std": null, "lo": null, "hi": null, "values": [0.707274490785645]},
    "auroc": {"n": 1, "mean": 0.8240956033135985, "std": null, "lo": null, "hi": null, "values": [0.8240956033135985]},
    "accuracy": {"n": 1, "mean": 0.7950841933731667, "std": null, "lo": null, "hi": null, "values": [0.7950841933731667]},
    "mcc": {"n": 1, "mean": 0.5496759487900071, "std": null, "lo": null, "hi": null, "values": [0.5496759487900071]},
    "precision": {"n": 1, "mean": 0.7115534738485558, "std": null, "lo": null, "hi": null, "values": [0.7115534738485558]},
    "recall": {"n": 1, "mean": 0.7030466640956421, "std": null, "lo": null, "hi": null, "values": [0.7030466640956421]},
    "ece": {"n": 1, "mean": 0.0643523095063318, "std": null, "lo": null, "hi": null, "values": [0.0643523095063318]},
    "adaptive_ece": {"n": 1, "mean": 0.0650151150016795, "std": null, "lo": null, "hi": null, "values": [0.0650151150016795]},
    "mce": {"n": 1, "mean": 0.107552745587523, "std": null, "lo": null, "hi": null, "values": [0.107552745587523]},
    "brier": {"n": 1, "mean": 0.3167558157567867, "std": null, "lo": null, "hi": null, "values": [0.3167558157567867]},
    "temperature": {"n": 1, "mean": 2.203227589733519, "std": null, "lo": null, "hi": null, "values": [2.203227589733519]},
    "ece_temp": {"n": 1, "mean": 0.1655349182277288, "std": null, "lo": null, "hi": null, "values": [0.1655349182277288]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.1655349182277287, "std": null, "lo": null, "hi": null, "values": [0.1655349182277287]},
    "mce_temp": {"n": 1, "mean": 0.1969301778144248, "std": null, "lo": null, "hi": null, "values": [0.1969301778144248]},
    "brier_temp": {"n": 1, "mean": 0.3646530045515972, "std": null, "lo": null, "hi": null, "values": [0.3646530045515972]}
   }},
  {"model": "gru", "task": "mitbih-afdb-binary", "averaging": "binary", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.9829378309472444, "std": null, "lo": null, "hi": null, "values": [0.9829378309472444]},
    "auroc": {"n": 1, "mean": 0.9992513258411492, "std": null, "lo": null, "hi": null, "values": [0.9992513258411492]},
    "accuracy": {"n": 1, "mean": 0.9881857686040196, "std": null, "lo": null, "hi": null, "values": [0.9881857686040196]},
    "mcc": {"n": 1, "mean": 0.9742383652827116, "std": null, "lo": null, "hi": null, "values": [0.9742383652827116]},
    "precision": {"n": 1, "mean": 1.0, "std": null, "lo": null, "hi": null, "values": [1.0]},
    "recall": {"n": 1, "mean": 0.9664481295796374, "std": null, "lo": null, "hi": null, "values": [0.9664481295796374]},
    "ece": {"n": 1, "mean": 0.0070193464230869, "std": null, "lo": null, "hi": null, "values": [0.0070193464230869]},
    "adaptive_ece": {"n": 1, "mean": 0.0067932004921466, "std": null, "lo": null, "hi": null, "values": [0.0067932004921466]},
    "mce": {"n": 1, "mean": 0.4634937246640523, "std": null, "lo": null, "hi": null, "values": [0.4634937246640523]},
    "brier": {"n": 1, "mean": 0.0192994205188456, "std": null, "lo": null, "hi": null, "values": [0.0192994205188456]},
    "temperature": {"n": 1, "mean": 2.142292446768524, "std": null, "lo": null, "hi": null, "values": [2.142292446768524]},
    "ece_temp": {"n": 1, "mean": 0.023409136676564, "std": null, "lo": null, "hi": null, "values": [0.023409136676564]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0205260953513813, "std": null, "lo": null, "hi": null, "values": [0.0205260953513813]},
    "mce_temp": {"n": 1, "mean": 0.2287137983870226, "std": null, "lo": null, "hi": null, "values": [0.2287137983870226]},
    "brier_temp": {"n": 1, "mean": 0.0196054216818733, "std": null, "lo": null, "hi": null, "values": [0.0196054216818733]}
   }},
  {"model": "inception-time", "task": "mitbih-afdb-binary", "averaging": "binary", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.8822700035248502, "std": null, "lo": null, "hi": null, "values": [0.8822700035248502]},
    "auroc": {"n": 1, "mean": 0.978602727640958, "std": null, "lo": null, "hi": null, "values": [0.978602727640958]},
    "accuracy": {"n": 1, "mean": 0.9092884302009776, "std": null, "lo": null, "hi": null, "values": [0.9092884302009776]},
    "mcc": {"n": 1, "mean": 0.8173383535114167, "std": null, "lo": null, "hi": null, "values": [0.8173383535114167]},
    "precision": {"n": 1, "mean": 0.8123985718922427, "std": null, "lo": null, "hi": null, "values": [0.8123985718922427]},
    "recall": {"n": 1, "mean": 0.9652911685306594, "std": null, "lo": null, "hi": null, "values": [0.9652911685306594]},
    "ece": {"n": 1, "mean": 0.0322701940469803, "std": null, "lo": null, "hi": null, "values": [0.0322701940469803]},
    "adaptive_ece": {"n": 1, "mean": 0.0301026092039897, "std": null, "lo": null, "hi": null, "values": [0.0301026092039897]},
    "mce": {"n": 1, "mean": 0.0615536802210788, "std": null, "lo": null, "hi": null, "values": [0.0615536802210788]},
    "brier": {"n": 1, "mean": 0.1439266356828333, "std": null, "lo": null, "hi": null, "values": [0.1439266356828333]},
    "temperature": {"n": 1, "mean": 1.558411546171708, "std": null, "lo": null, "hi": null, "values": [1.558411546171708]},
    "ece_temp": {"n": 1, "mean": 0.0206609490339043, "std": null, "lo": null, "hi": null, "values": [0.0206609490339043]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0208375487463757, "std": null, "lo": null, "hi": null, "values": [0.0208375487463757]},
    "mce_temp": {"n": 1, "mean": 0.074160392697363, "std": null, "lo": null, "hi": null, "values": [0.074160392697363]},
    "brier_temp": {"n": 1, "mean": 0.1428628916099767, "std": null, "lo": null, "hi": null, "values": [0.1428628916099767]}
   }},
  {"model": "kanres", "task": "mitbih-afdb-binary", "averaging": "binary", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.3658385093167702, "std": null, "lo": null, "hi": null, "values": [0.3658385093167702]},
    "auroc": {"n": 1, "mean": 0.457126845303565, "std": null, "lo": null, "hi": null, "values": [0.457126845303565]},
    "accuracy": {"n": 1, "mean": 0.44541010320478, "std": null, "lo": null, "hi": null, "values": [0.44541010320478]},
    "mcc": {"n": 1, "mean": -0.1005194490786472, "std": null, "lo": null, "hi": null, "values": [-0.1005194490786472]},
    "precision": {"n": 1, "mean": 0.3062126332206914, "std": null, "lo": null, "hi": null, "values": [0.3062126332206914]},
    "recall": {"n": 1, "mean": 0.4543000385653683, "std": null, "lo": null, "hi": null, "values": [0.4543000385653683]},
    "ece": {"n": 1, "mean": 0.1976155144066865, "std": null, "lo": null, "hi": null, "values": [0.1976155144066865]},
    "adaptive_ece": {"n": 1, "mean": 0.1976155144066865, "std": null, "lo": null, "hi": null, "values": [0.1976155144066865]},
    "mce": {"n": 1, "mean": 0.9110830163955688, "std": null, "lo": null, "hi": null, "values": [0.9110830163955688]},
    "brier": {"n": 1, "mean": 0.5857219590181983, "std": null, "lo": null, "hi": null, "values": [0.5857219590181983]},
    "temperature": {"n": 1, "mean": 10.711551025410834, "std": null, "lo": null, "hi": null, "values": [10.711551025410834]},
    "ece_temp": {"n": 1, "mean": 0.0693209237150275, "std": null, "lo": null, "hi": null, "values": [0.0693209237150275]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0741431340446037, "std": null, "lo": null, "hi": null, "values": [0.0741431340446037]},
    "mce_temp": {"n": 1, "mean": 0.1388576492011723, "std": null, "lo": null, "hi": null, "values": [0.1388576492011723]},
    "brier_temp": {"n": 1, "mean": 0.5035138433937222, "std": null, "lo": null, "hi": null, "values": [0.5035138433937222]}
   }},
  {"model": "kanres-deep", "task": "mitbih-afdb-binary", "averaging": "binary", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.9797191887675508, "std": null, "lo": null, "hi": null, "values": [0.9797191887675508]},
    "auroc": {"n": 1, "mean": 0.9971285735105956, "std": null, "lo": null, "hi": null, "values": [0.9971285735105956]},
    "accuracy": {"n": 1, "mean": 0.9858772406300924, "std": null, "lo": null, "hi": null, "values": [0.9858772406300924]},
    "mcc": {"n": 1, "mean": 0.9690338586136882, "std": null, "lo": null, "hi": null, "values": [0.9690338586136882]},
    "precision": {"n": 1, "mean": 0.9909270216962524, "std": null, "lo": null, "hi": null, "values": [0.9909270216962524]},
    "recall": {"n": 1, "mean": 0.9687620516775935, "std": null, "lo": null, "hi": null, "values": [0.9687620516775935]},
    "ece": {"n": 1, "mean": 0.0091132990437186, "std": null, "lo": null, "hi": null, "values": [0.0091132990437186]},
    "adaptive_ece": {"n": 1, "mean": 0.0136404642263616, "std": null, "lo": null, "hi": null, "values": [0.0136404642263616]},
    "mce": {"n": 1, "mean": 0.1900348591804505, "std": null, "lo": null, "hi": null, "values": [0.1900348591804505]},
    "brier": {"n": 1, "mean": 0.0265407228465259, "std": null, "lo": null, "hi": null, "values": [0.0265407228465259]},
    "temperature": {"n": 1, "mean": 2.6781000321785853, "std": null, "lo": null, "hi": null, "values": [2.6781000321785853]},
    "ece_temp": {"n": 1, "mean": 0.0883009188234058, "std": null, "lo": null, "hi": null, "values": [0.0883009188234058]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0883009188234059, "std": null, "lo": null, "hi": null, "values": [0.0883009188234059]},
    "mce_temp": {"n": 1, "mean": 0.314317630515785, "std": null, "lo": null, "hi": null, "values": [0.314317630515785]},
    "brier_temp": {"n": 1, "mean": 0.0482637807287372, "std": null, "lo": null, "hi": null, "values": [0.0482637807287372]}
   }},
  {"model": "lstm", "task": "mitbih-afdb-binary", "averaging": "binary", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.0, "std": null, "lo": null, "hi": null, "values": [0.0]},
    "auroc": {"n": 1, "mean": 0.4580361748166286, "std": null, "lo": null, "hi": null, "values": [0.4580361748166286]},
    "accuracy": {"n": 1, "mean": 0.6478815860945139, "std": null, "lo": null, "hi": null, "values": [0.6478815860945139]},
    "mcc": {"n": 1, "mean": 0.0, "std": null, "lo": null, "hi": null, "values": [0.0]},
    "precision": {"n": 1, "mean": 0.0, "std": null, "lo": null, "hi": null, "values": [0.0]},
    "recall": {"n": 1, "mean": 0.0, "std": null, "lo": null, "hi": null, "values": [0.0]},
    "ece": {"n": 1, "mean": 0.0504186515863791, "std": null, "lo": null, "hi": null, "values": [0.0504186515863791]},
    "adaptive_ece": {"n": 1, "mean": 0.0622924427536762, "std": null, "lo": null, "hi": null, "values": [0.0622924427536762]},
    "mce": {"n": 1, "mean": 0.0508978628492545, "std": null, "lo": null, "hi": null, "values": [0.0508978628492545]},
    "brier": {"n": 1, "mean": 0.4613441322553592, "std": null, "lo": null, "hi": null, "values": [0.4613441322553592]},
    "temperature": {"n": 1, "mean": 1.2590125343046568, "std": null, "lo": null, "hi": null, "values": [1.2590125343046568]},
    "ece_temp": {"n": 1, "mean": 0.0701020675760338, "std": null, "lo": null, "hi": null, "values": [0.0701020675760338]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0743896389105586, "std": null, "lo": null, "hi": null, "values": [0.0743896389105586]},
    "mce_temp": {"n": 1, "mean": 0.2534280233142635, "std": null, "lo": null, "hi": null, "values": [0.2534280233142635]},
    "brier_temp": {"n": 1, "mean": 0.4660876668969989, "std": null, "lo": null, "hi": null, "values": [0.4660876668969989]}
   }},
  {"model": "medformer", "task": "mitbih-afdb-binary", "averaging": "binary", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.912770137524558, "std": null, "lo": null, "hi": null, "values": [0.912770137524558]},
    "auroc": {"n": 1, "mean": 0.9757748700752868, "std": null, "lo": null, "hi": null, "values": [0.9757748700752868]},
    "accuracy": {"n": 1, "mean": 0.939706681151548, "std": null, "lo": null, "hi": null, "values": [0.939706681151548]},
    "mcc": {"n": 1, "mean": 0.8670875901497694, "std": null, "lo": null, "hi": null, "values": [0.8670875901497694]},
    "precision": {"n": 1, "mean": 0.9303163796555868, "std": null, "lo": null, "hi": null, "values": [0.9303163796555868]},
    "recall": {"n": 1, "mean": 0.8958735055919784, "std": null, "lo": null, "hi": null, "values": [0.8958735055919784]},
    "ece": {"n": 1, "mean": 0.0122720208959952, "std": null, "lo": null, "hi": null, "values": [0.0122720208959952]},
    "adaptive_ece": {"n": 1, "mean": 0.0164535238667068, "std": null, "lo": null, "hi": null, "values": [0.0164535238667068]},
    "mce": {"n": 1, "mean": 0.0856087355203525, "std": null, "lo": null, "hi": null, "values": [0.0856087355203525]},
    "brier": {"n": 1, "mean": 0.0959115746647802, "std": null, "lo": null, "hi": null, "values": [0.0959115746647802]},
    "temperature": {"n": 1, "mean": 3.17921125731616, "std": null, "lo": null, "hi": null, "values": [3.17921125731616]},
    "ece_temp": {"n": 1, "mean": 0.1839561395125478, "std": null, "lo": null, "hi": null, "values": [0.1839561395125478]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.1839561395125477, "std": null, "lo": null, "hi": null, "values": [0.1839561395125477]},
    "mce_temp": {"n": 1, "mean": 0.2610109532320128, "std": null, "lo": null, "hi": null, "values": [0.2610109532320128]},
    "brier_temp": {"n": 1, "mean": 0.1704794714710941, "std": null, "lo": null, "hi": null, "values": [0.1704794714710941]}
   }},
  {"model": "mlp-wang", "task": "mitbih-afdb-binary", "averaging": "binary", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.6081734778982485, "std": null, "lo": null, "hi": null, "values": [0.6081734778982485]},
    "auroc": {"n": 1, "mean": 0.7507422681529032, "std": null, "lo": null, "hi": null, "values": [0.7507422681529032]},
    "accuracy": {"n": 1, "mean": 0.681015752308528, "std": null, "lo": null, "hi": null, "values": [0.681015752308528]},
    "mcc": {"n": 1, "mean": 0.3564742523271539, "std": null, "lo": null, "hi": null, "values": [0.3564742523271539]},
    "precision": {"n": 1, "mean": 0.5358612580834803, "std": null, "lo": null, "hi": null, "values": [0.5358612580834803]},
    "recall": {"n": 1, "mean": 0.7030466640956421, "std": null, "lo": null, "hi": null, "values": [0.7030466640956421]},
    "ece": {"n": 1, "mean": 0.0557836139739563, "std": null, "lo": null, "hi": null, "values": [0.0557836139739563]},
    "adaptive_ece": {"n": 1, "mean": 0.0577960427949637, "std": null, "lo": null, "hi": null, "values": [0.0577960427949637]},
    "mce": {"n": 1, "mean": 0.0909821671393987, "std": null, "lo": null, "hi": null, "values": [0.0909821671393987]},
    "brier": {"n": 1, "mean": 0.4150014670004338, "std": null, "lo": null, "hi": null, "values": [0.4150014670004338]},
    "temperature": {"n": 1, "mean": 1.9799354251386936, "std": null, "lo": null, "hi": null, "values": [1.9799354251386936]},
    "ece_temp": {"n": 1, "mean": 0.1118710623649514, "std": null, "lo": null, "hi": null, "values": [0.1118710623649514]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.1118812742906123, "std": null, "lo": null, "hi": null, "values": [0.1118812742906123]},
    "mce_temp": {"n": 1, "mean": 0.1912834780630437, "std": null, "lo": null, "hi": null, "values": [0.1912834780630437]},
    "brier_temp": {"n": 1, "mean": 0.4391488995976974, "std": null, "lo": null, "hi": null, "values": [0.4391488995976974]}
   }},
  {"model": "resnet", "task": "mitbih-afdb-binary", "averaging": "binary", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.980422866092404, "std": null, "lo": null, "hi": null, "values": [0.980422866092404]},
    "auroc": {"n": 1, "mean": 0.9990458486535222, "std": null, "lo": null, "hi": null, "values": [0.9990458486535222]},
    "accuracy": {"n": 1, "mean": 0.986420423682781, "std": null, "lo": null, "hi": null, "values": [0.986420423682781]},
    "mcc": {"n": 1, "mean": 0.9702966914653536, "std": null, "lo": null, "hi": null, "values": [0.9702966914653536]},
    "precision": {"n": 1, "mean": 0.9956262425447316, "std": null, "lo": null, "hi": null, "values": [0.9956262425447316]},
    "recall": {"n": 1, "mean": 0.965676822213652, "std": null, "lo": null, "hi": null, "values": [0.965676822213652]},
    "ece": {"n": 1, "mean": 0.0083774108073427, "std": null, "lo": null, "hi": null, "values": [0.0083774108073427]},
    "adaptive_ece": {"n": 1, "mean": 0.008963078988887, "std": null, "lo": null, "hi": null, "values": [0.008963078988887]},
    "mce": {"n": 1, "mean": 0.5147927304108938, "std": null, "lo": null, "hi": null, "values": [0.5147927304108938]},
    "brier": {"n": 1, "mean": 0.0222184652537825, "std": null, "lo": null, "hi": null, "values": [0.0222184652537825]},
    "temperature": {"n": 1, "mean": 2.748016970015028, "std": null, "lo": null, "hi": null, "values": [2.748016970015028]},
    "ece_temp": {"n": 1, "mean": 0.0496798708069119, "std": null, "lo": null, "hi": null, "values": [0.0496798708069119]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0481704524667192, "std": null, "lo": null, "hi": null, "values": [0.0481704524667192]},
    "mce_temp": {"n": 1, "mean": 0.2502563904940756, "std": null, "lo": null, "hi": null, "values": [0.2502563904940756]},
    "brier_temp": {"n": 1, "mean": 0.0272656256552493, "std": null, "lo": null, "hi": null, "values": [0.0272656256552493]}
   }},
  {"model": "resnet-hannun", "task": "mitbih-afdb-binary", "averaging": "binary", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.9827788649706458, "std": null, "lo": null, "hi": null, "values": [0.9827788649706458]},
    "auroc": {"n": 1, "mean": 0.9996330995457756, "std": null, "lo": null, "hi": null, "values": [0.9996330995457756]},
    "accuracy": {"n": 1, "mean": 0.9880499728408474, "std": null, "lo": null, "hi": null, "values": [0.9880499728408474]},
    "mcc": {"n": 1, "mean": 0.9738849954062078, "std": null, "lo": null, "hi": null, "values": [0.9738849954062078]},
    "precision": {"n": 1, "mean": 0.99761620977354, "std": null, "lo": null, "hi": null, "values": [0.99761620977354]},
    "recall": {"n": 1, "mean": 0.9683763979946008, "std": null, "lo": null, "hi": null, "values": [0.9683763979946008]},
    "ece": {"n": 1, "mean": 0.0065224225481588, "std": null, "lo": null, "hi": null, "values": [0.0065224225481588]},
    "adaptive_ece": {"n": 1, "mean": 0.0068401190604547, "std": null, "lo": null, "hi": null, "values": [0.0068401190604547]},
    "mce": {"n": 1, "mean": 0.4750472095277574, "std": null, "lo": null, "hi": null, "values": [0.4750472095277574]},
    "brier": {"n": 1, "mean": 0.0173728294101792, "std": null, "lo": null, "hi": null, "values": [0.0173728294101792]},
    "temperature": {"n": 1, "mean": 2.3736072859081574, "std": null, "lo": null, "hi": null, "values": [2.3736072859081574]},
    "ece_temp": {"n": 1, "mean": 0.0271118338623593, "std": null, "lo": null, "hi": null, "values": [0.0271118338623593]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0214347071270599, "std": null, "lo": null, "hi": null, "values": [0.0214347071270599]},
    "mce_temp": {"n": 1, "mean": 0.3168727196030181, "std": null, "lo": null, "hi": null, "values": [0.3168727196030181]},
    "brier_temp": {"n": 1, "mean": 0.0179643954122285, "std": null, "lo": null, "hi": null, "values": [0.0179643954122285]}
   }},
  {"model": "resnet-wang", "task": "mitbih-afdb-binary", "averaging": "binary", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.4933675652545999, "std": null, "lo": null, "hi": null, "values": [0.4933675652545999]},
    "auroc": {"n": 1, "mean": 0.7287708398285924, "std": null, "lo": null, "hi": null, "values": [0.7287708398285924]},
    "accuracy": {"n": 1, "mean": 0.6784356328082564, "std": null, "lo": null, "hi": null, "values": [0.6784356328082564]},
    "mcc": {"n": 1, "mean": 0.2653569635989064, "std": null, "lo": null, "hi": null, "values": [0.2653569635989064]},
    "precision": {"n": 1, "mean": 0.5540605478135512, "std": null, "lo": null, "hi": null, "values": [0.5540605478135512]},
    "recall": {"n": 1, "mean": 0.4446586964905515, "std": null, "lo": null, "hi": null, "values": [0.4446586964905515]},
    "ece": {"n": 1, "mean": 0.1643929630922662, "std": null, "lo": null, "hi": null, "values": [0.1643929630922662]},
    "adaptive_ece": {"n": 1, "mean": 0.1643929630922662, "std": null, "lo": null, "hi": null, "values": [0.1643929630922662]},
    "mce": {"n": 1, "mean": 0.2564327898057731, "std": null, "lo": null, "hi": null, "values": [0.2564327898057731]},
    "brier": {"n": 1, "mean": 0.4672782345950886, "std": null, "lo": null, "hi": null, "values": [0.4672782345950886]},
    "temperature": {"n": 1, "mean": 2.8367057103991424, "std": null, "lo": null, "hi": null, "values": [2.8367057103991424]},
    "ece_temp": {"n": 1, "mean": 0.0238127248936356, "std": null, "lo": null, "hi": null, "values": [0.0238127248936356]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.025709167156121, "std": null, "lo": null, "hi": null, "values": [0.025709167156121]},
    "mce_temp": {"n": 1, "mean": 0.0577811848612458, "std": null, "lo": null, "hi": null, "values": [0.0577811848612458]},
    "brier_temp": {"n": 1, "mean": 0.4023789449818689, "std": null, "lo": null, "hi": null, "values": [0.4023789449818689]}
   }},
  {"model": "rocket", "task": "mitbih-afdb-binary", "averaging": "binary", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.9296567248171076, "std": null, "lo": null, "hi": null, "values": [0.9296567248171076]},
    "auroc": {"n": 1, "mean": 0.9864678075365833, "std": null, "lo": null, "hi": null, "values": [0.9864678075365833]},
    "accuracy": {"n": 1, "mean": 0.9490765888104292, "std": null, "lo": null, "hi": null, "values": [0.9490765888104292]},
    "mcc": {"n": 1, "mean": 0.8906055633276648, "std": null, "lo": null, "hi": null, "values": [0.8906055633276648]},
    "precision": {"n": 1, "mean": 0.9050401753104456, "std": null, "lo": null, "hi": null, "values": [0.9050401753104456]},
    "recall": {"n": 1, "mean": 0.9556498264558426, "std": null, "lo": null, "hi": null, "values": [0.9556498264558426]},
    "ece": {"n": 1, "mean": 0.0566972198854122, "std": null, "lo": null, "hi": null, "values": [0.0566972198854122]},
    "adaptive_ece": {"n": 1, "mean": 0.0567664800293739, "std": null, "lo": null, "hi": null, "values": [0.0567664800293739]},
    "mce": {"n": 1, "mean": 0.1413953309994564, "std": null, "lo": null, "hi": null, "values": [0.1413953309994564]},
    "brier": {"n": 1, "mean": 0.090421599411497, "std": null, "lo": null, "hi": null, "values": [0.090421599411497]},
    "temperature": {"n": 1, "mean": 1.5843237108023271, "std": null, "lo": null, "hi": null, "values": [1.5843237108023271]},
    "ece_temp": {"n": 1, "mean": 0.1231391337500772, "std": null, "lo": null, "hi": null, "values": [0.1231391337500772]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.1231391337500772, "std": null, "lo": null, "hi": null, "values": [0.1231391337500772]},
    "mce_temp": {"n": 1, "mean": 0.2205453505853055, "std": null, "lo": null, "hi": null, "values": [0.2205453505853055]},
    "brier_temp": {"n": 1, "mean": 0.1193767753023139, "std": null, "lo": null, "hi": null, "values": [0.1193767753023139]}
   }},
  {"model": "se-resnet", "task": "mitbih-afdb-binary", "averaging": "binary", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.9817825661116552, "std": null, "lo": null, "hi": null, "values": [0.9817825661116552]},
    "auroc": {"n": 1, "mean": 0.9934874563128582, "std": null, "lo": null, "hi": null, "values": [0.9934874563128582]},
    "accuracy": {"n": 1, "mean": 0.9873709940249864, "std": null, "lo": null, "hi": null, "values": [0.9873709940249864]},
    "mcc": {"n": 1, "mean": 0.9724086331740448, "std": null, "lo": null, "hi": null, "values": [0.9724086331740448]},
    "precision": {"n": 1, "mean": 0.9976114649681528, "std": null, "lo": null, "hi": null, "values": [0.9976114649681528]},
    "recall": {"n": 1, "mean": 0.9664481295796374, "std": null, "lo": null, "hi": null, "values": [0.9664481295796374]},
    "ece": {"n": 1, "mean": 0.0108649719107869, "std": null, "lo": null, "hi": null, "values": [0.0108649719107869]},
    "adaptive_ece": {"n": 1, "mean": 0.009871895281858, "std": null, "lo": null, "hi": null, "values": [0.009871895281858]},
    "mce": {"n": 1, "mean": 0.1671854456265767, "std": null, "lo": null, "hi": null, "values": [0.1671854456265767]},
    "brier": {"n": 1, "mean": 0.0243343715954276, "std": null, "lo": null, "hi": null, "values": [0.0243343715954276]},
    "temperature": {"n": 1, "mean": 2.594315906108244, "std": null, "lo": null, "hi": null, "values": [2.594315906108244]},
    "ece_temp": {"n": 1, "mean": 0.0186511606512097, "std": null, "lo": null, "hi": null, "values": [0.0186511606512097]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0184979179087813, "std": null, "lo": null, "hi": null, "values": [0.0184979179087813]},
    "mce_temp": {"n": 1, "mean": 0.3200143683127282, "std": null, "lo": null, "hi": null, "values": [0.3200143683127282]},
    "brier_temp": {"n": 1, "mean": 0.0262731680488282, "std": null, "lo": null, "hi": null, "values": [0.0262731680488282]}
   }},
  {"model": "simple-cnn", "task": "mitbih-afdb-binary", "averaging": "binary", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.9001463057790783, "std": null, "lo": null, "hi": null, "values": [0.9001463057790783]},
    "auroc": {"n": 1, "mean": 0.9707427806333788, "std": null, "lo": null, "hi": null, "values": [0.9707427806333788]},
    "accuracy": {"n": 1, "mean": 0.9258555133079848, "std": null, "lo": null, "hi": null, "values": [0.9258555133079848]},
    "mcc": {"n": 1, "mean": 0.8442681828301746, "std": null, "lo": null, "hi": null, "values": [0.8442681828301746]},
    "precision": {"n": 1, "mean": 0.856, "std": null, "lo": null, "hi": null, "values": [0.856]},
    "recall": {"n": 1, "mean": 0.9490937138449672, "std": null, "lo": null, "hi": null, "values": [0.9490937138449672]},
    "ece": {"n": 1, "mean": 0.0294301986192503, "std": null, "lo": null, "hi": null, "values": [0.0294301986192503]},
    "adaptive_ece": {"n": 1, "mean": 0.0337525720038794, "std": null, "lo": null, "hi": null, "values": [0.0337525720038794]},
    "mce": {"n": 1, "mean": 0.1053794414076761, "std": null, "lo": null, "hi": null, "values": [0.1053794414076761]},
    "brier": {"n": 1, "mean": 0.1191757367070564, "std": null, "lo": null, "hi": null, "values": [0.1191757367070564]},
    "temperature": {"n": 1, "mean": 3.99811034521101, "std": null, "lo": null, "hi": null, "values": [3.99811034521101]},
    "ece_temp": {"n": 1, "mean": 0.2163814695156564, "std": null, "lo": null, "hi": null, "values": [0.2163814695156564]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.2163814695156564, "std": null, "lo": null, "hi": null, "values": [0.2163814695156564]},
    "mce_temp": {"n": 1, "mean": 0.3150550690713676, "std": null, "lo": null, "hi": null, "values": [0.3150550690713676]},
    "brier_temp": {"n": 1, "mean": 0.2228543898797305, "std": null, "lo": null, "hi": null, "values": [0.2228543898797305]}
   }},
  {"model": "spectrogram-cnn", "task": "mitbih-afdb-binary", "averaging": "binary", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.9713617767387492, "std": null, "lo": null, "hi": null, "values": [0.9713617767387492]},
    "auroc": {"n": 1, "mean": 0.997229372115226, "std": null, "lo": null, "hi": null, "values": [0.997229372115226]},
    "accuracy": {"n": 1, "mean": 0.9800380228136882, "std": null, "lo": null, "hi": null, "values": [0.9800380228136882]},
    "mcc": {"n": 1, "mean": 0.9561638888091956, "std": null, "lo": null, "hi": null, "values": [0.9561638888091956]},
    "precision": {"n": 1, "mean": 0.981496062992126, "std": null, "lo": null, "hi": null, "values": [0.981496062992126]},
    "recall": {"n": 1, "mean": 0.9614346317007328, "std": null, "lo": null, "hi": null, "values": [0.9614346317007328]},
    "ece": {"n": 1, "mean": 0.0123108685874602, "std": null, "lo": null, "hi": null, "values": [0.0123108685874602]},
    "adaptive_ece": {"n": 1, "mean": 0.0118556744229333, "std": null, "lo": null, "hi": null, "values": [0.0118556744229333]},
    "mce": {"n": 1, "mean": 0.1516796725136893, "std": null, "lo": null, "hi": null, "values": [0.1516796725136893]},
    "brier": {"n": 1, "mean": 0.036415329800599, "std": null, "lo": null, "hi": null, "values": [0.036415329800599]},
    "temperature": {"n": 1, "mean": 1.70268573337176, "std": null, "lo": null, "hi": null, "values": [1.70268573337176]},
    "ece_temp": {"n": 1, "mean": 0.0181401332733465, "std": null, "lo": null, "hi": null, "values": [0.0181401332733465]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0169626242589191, "std": null, "lo": null, "hi": null, "values": [0.0169626242589191]},
    "mce_temp": {"n": 1, "mean": 0.2438039646612959, "std": null, "lo": null, "hi": null, "values": [0.2438039646612959]},
    "brier_temp": {"n": 1, "mean": 0.0379007362561395, "std": null, "lo": null, "hi": null, "values": [0.0379007362561395]}
   }},
  {"model": "tcn", "task": "mitbih-afdb-binary", "averaging": "binary", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.9813542688910696, "std": null, "lo": null, "hi": null, "values": [0.9813542688910696]},
    "auroc": {"n": 1, "mean": 0.9994865495295808, "std": null, "lo": null, "hi": null, "values": [0.9994865495295808]},
    "accuracy": {"n": 1, "mean": 0.987099402498642, "std": null, "lo": null, "hi": null, "values": [0.987099402498642]},
    "mcc": {"n": 1, "mean": 0.9718594437233812, "std": null, "lo": null, "hi": null, "values": [0.9718594437233812]},
    "precision": {"n": 1, "mean": 0.9992006394884092, "std": null, "lo": null, "hi": null, "values": [0.9992006394884092]},
    "recall": {"n": 1, "mean": 0.9641342074816814, "std": null, "lo": null, "hi": null, "values": [0.9641342074816814]},
    "ece": {"n": 1, "mean": 0.0120515195912864, "std": null, "lo": null, "hi": null, "values": [0.0120515195912864]},
    "adaptive_ece": {"n": 1, "mean": 0.0118562732456431, "std": null, "lo": null, "hi": null, "values": [0.0118562732456431]},
    "mce": {"n": 1, "mean": 0.3690683841705322, "std": null, "lo": null, "hi": null, "values": [0.3690683841705322]},
    "brier": {"n": 1, "mean": 0.024890758319827, "std": null, "lo": null, "hi": null, "values": [0.024890758319827]},
    "temperature": {"n": 1, "mean": 2.0852088981552876, "std": null, "lo": null, "hi": null, "values": [2.0852088981552876]},
    "ece_temp": {"n": 1, "mean": 0.007323867650784, "std": null, "lo": null, "hi": null, "values": [0.007323867650784]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0085910372686889, "std": null, "lo": null, "hi": null, "values": [0.0085910372686889]},
    "mce_temp": {"n": 1, "mean": 0.3641047491242667, "std": null, "lo": null, "hi": null, "values": [0.3641047491242667]},
    "brier_temp": {"n": 1, "mean": 0.0216920942888858, "std": null, "lo": null, "hi": null, "values": [0.0216920942888858]}
   }},
  {"model": "xresnet", "task": "mitbih-afdb-binary", "averaging": "binary", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.9791420700511608, "std": null, "lo": null, "hi": null, "values": [0.9791420700511608]},
    "auroc": {"n": 1, "mean": 0.9998123868794327, "std": null, "lo": null, "hi": null, "values": [0.9998123868794327]},
    "accuracy": {"n": 1, "mean": 0.985605649103748, "std": null, "lo": null, "hi": null, "values": [0.985605649103748]},
    "mcc": {"n": 1, "mean": 0.9686328368605552, "std": null, "lo": null, "hi": null, "values": [0.9686328368605552]},
    "precision": {"n": 1, "mean": 0.9995982322217758, "std": null, "lo": null, "hi": null, "values": [0.9995982322217758]},
    "recall": {"n": 1, "mean": 0.9595063632857692, "std": null, "lo": null, "hi": null, "values": [0.9595063632857692]},
    "ece": {"n": 1, "mean": 0.0085911143496268, "std": null, "lo": null, "hi": null, "values": [0.0085911143496268]},
    "adaptive_ece": {"n": 1, "mean": 0.0089108535917733, "std": null, "lo": null, "hi": null, "values": [0.0089108535917733]},
    "mce": {"n": 1, "mean": 0.3202927336096763, "std": null, "lo": null, "hi": null, "values": [0.3202927336096763]},
    "brier": {"n": 1, "mean": 0.0244340608828828, "std": null, "lo": null, "hi": null, "values": [0.0244340608828828]},
    "temperature": {"n": 1, "mean": 2.177071973008669, "std": null, "lo": null, "hi": null, "values": [2.177071973008669]},
    "ece_temp": {"n": 1, "mean": 0.0338945851334189, "std": null, "lo": null, "hi": null, "values": [0.0338945851334189]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0267607257202828, "std": null, "lo": null, "hi": null, "values": [0.0267607257202828]},
    "mce_temp": {"n": 1, "mean": 0.1983518146571634, "std": null, "lo": null, "hi": null, "values": [0.1983518146571634]},
    "brier_temp": {"n": 1, "mean": 0.0248258573611016, "std": null, "lo": null, "hi": null, "values": [0.0248258573611016]}
   }},
  {"model": "xresnet1d-benchmark", "task": "mitbih-afdb-binary", "averaging": "binary", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.9797523098093178, "std": null, "lo": null, "hi": null, "values": [0.9797523098093178]},
    "auroc": {"n": 1, "mean": 0.9740271823201024, "std": null, "lo": null, "hi": null, "values": [0.9740271823201024]},
    "accuracy": {"n": 1, "mean": 0.9860130363932644, "std": null, "lo": null, "hi": null, "values": [0.9860130363932644]},
    "mcc": {"n": 1, "mean": 0.9695033336592952, "std": null, "lo": null, "hi": null, "values": [0.9695033336592952]},
    "precision": {"n": 1, "mean": 0.9991980753809142, "std": null, "lo": null, "hi": null, "values": [0.9991980753809142]},
    "recall": {"n": 1, "mean": 0.96104897801774, "std": null, "lo": null, "hi": null, "values": [0.96104897801774]},
    "ece": {"n": 1, "mean": 0.00974027489528, "std": null, "lo": null, "hi": null, "values": [0.00974027489528]},
    "adaptive_ece": {"n": 1, "mean": 0.0115794226535308, "std": null, "lo": null, "hi": null, "values": [0.0115794226535308]},
    "mce": {"n": 1, "mean": 0.1662456132471561, "std": null, "lo": null, "hi": null, "values": [0.1662456132471561]},
    "brier": {"n": 1, "mean": 0.0269956600384067, "std": null, "lo": null, "hi": null, "values": [0.0269956600384067]},
    "temperature": {"n": 1, "mean": 1.529707768513325, "std": null, "lo": null, "hi": null, "values": [1.529707768513325]},
    "ece_temp": {"n": 1, "mean": 0.0060061319115315, "std": null, "lo": null, "hi": null, "values": [0.0060061319115315]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0228856356589371, "std": null, "lo": null, "hi": null, "values": [0.0228856356589371]},
    "mce_temp": {"n": 1, "mean": 0.1828779494115087, "std": null, "lo": null, "hi": null, "values": [0.1828779494115087]},
    "brier_temp": {"n": 1, "mean": 0.0284940053743928, "std": null, "lo": null, "hi": null, "values": [0.0284940053743928]}
   }},
  {"model": "xresnet1d101", "task": "mitbih-afdb-binary", "averaging": "binary", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.9769458128078816, "std": null, "lo": null, "hi": null, "values": [0.9769458128078816]},
    "auroc": {"n": 1, "mean": 0.985521537396161, "std": null, "lo": null, "hi": null, "values": [0.985521537396161]},
    "accuracy": {"n": 1, "mean": 0.984111895708854, "std": null, "lo": null, "hi": null, "values": [0.984111895708854]},
    "mcc": {"n": 1, "mean": 0.9653714804855666, "std": null, "lo": null, "hi": null, "values": [0.9653714804855666]},
    "precision": {"n": 1, "mean": 0.998791297340854, "std": null, "lo": null, "hi": null, "values": [0.998791297340854]},
    "recall": {"n": 1, "mean": 0.9560354801388352, "std": null, "lo": null, "hi": null, "values": [0.9560354801388352]},
    "ece": {"n": 1, "mean": 0.0119819057288861, "std": null, "lo": null, "hi": null, "values": [0.0119819057288861]},
    "adaptive_ece": {"n": 1, "mean": 0.014565582567748, "std": null, "lo": null, "hi": null, "values": [0.014565582567748]},
    "mce": {"n": 1, "mean": 0.1422034942585489, "std": null, "lo": null, "hi": null, "values": [0.1422034942585489]},
    "brier": {"n": 1, "mean": 0.0307069254906021, "std": null, "lo": null, "hi": null, "values": [0.0307069254906021]},
    "temperature": {"n": 1, "mean": 1.7316224923599313, "std": null, "lo": null, "hi": null, "values": [1.7316224923599313]},
    "ece_temp": {"n": 1, "mean": 0.0157900934011083, "std": null, "lo": null, "hi": null, "values": [0.0157900934011083]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0284303970940218, "std": null, "lo": null, "hi": null, "values": [0.0284303970940218]},
    "mce_temp": {"n": 1, "mean": 0.2354788959972467, "std": null, "lo": null, "hi": null, "values": [0.2354788959972467]},
    "brier_temp": {"n": 1, "mean": 0.0344490011722648, "std": null, "lo": null, "hi": null, "values": [0.0344490011722648]}
   }},
  {"model": "convnext-v2", "task": "ltafdb-binary", "averaging": "binary", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.821013081470502, "std": null, "lo": null, "hi": null, "values": [0.821013081470502]},
    "auroc": {"n": 1, "mean": 0.9405570633045748, "std": null, "lo": null, "hi": null, "values": [0.9405570633045748]},
    "accuracy": {"n": 1, "mean": 0.8430652522138604, "std": null, "lo": null, "hi": null, "values": [0.8430652522138604]},
    "mcc": {"n": 1, "mean": 0.7202432141403933, "std": null, "lo": null, "hi": null, "values": [0.7202432141403933]},
    "precision": {"n": 1, "mean": 0.9882167454982989, "std": null, "lo": null, "hi": null, "values": [0.9882167454982989]},
    "recall": {"n": 1, "mean": 0.7022023054924968, "std": null, "lo": null, "hi": null, "values": [0.7022023054924968]},
    "ece": {"n": 1, "mean": 0.1326230807921127, "std": null, "lo": null, "hi": null, "values": [0.1326230807921127]},
    "adaptive_ece": {"n": 1, "mean": 0.1326230807921127, "std": null, "lo": null, "hi": null, "values": [0.1326230807921127]},
    "mce": {"n": 1, "mean": 0.460107719030464, "std": null, "lo": null, "hi": null, "values": [0.460107719030464]},
    "brier": {"n": 1, "mean": 0.2766925959787801, "std": null, "lo": null, "hi": null, "values": [0.2766925959787801]},
    "temperature": {"n": 1, "mean": 2.498544485933317, "std": null, "lo": null, "hi": null, "values": [2.498544485933317]},
    "ece_temp": {"n": 1, "mean": 0.0580261280285436, "std": null, "lo": null, "hi": null, "values": [0.0580261280285436]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0815005985920595, "std": null, "lo": null, "hi": null, "values": [0.0815005985920595]},
    "mce_temp": {"n": 1, "mean": 0.383258799771822, "std": null, "lo": null, "hi": null, "values": [0.383258799771822]},
    "brier_temp": {"n": 1, "mean": 0.2249518331428262, "std": null, "lo": null, "hi": null, "values": [0.2249518331428262]}
   }},
  {"model": "crnn", "task": "ltafdb-binary", "averaging": "binary", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.861643111874257, "std": null, "lo": null, "hi": null, "values": [0.861643111874257]},
    "auroc": {"n": 1, "mean": 0.9706366557047336, "std": null, "lo": null, "hi": null, "values": [0.9706366557047336]},
    "accuracy": {"n": 1, "mean": 0.8733641611509052, "std": null, "lo": null, "hi": null, "values": [0.8733641611509052]},
    "mcc": {"n": 1, "mean": 0.7665091292566253, "std": null, "lo": null, "hi": null, "values": [0.7665091292566253]},
    "precision": {"n": 1, "mean": 0.979173702577958, "std": null, "lo": null, "hi": null, "values": [0.979173702577958]},
    "recall": {"n": 1, "mean": 0.7693033403107403, "std": null, "lo": null, "hi": null, "values": [0.7693033403107403]},
    "ece": {"n": 1, "mean": 0.0878848856853819, "std": null, "lo": null, "hi": null, "values": [0.0878848856853819]},
    "adaptive_ece": {"n": 1, "mean": 0.087884885685382, "std": null, "lo": null, "hi": null, "values": [0.087884885685382]},
    "mce": {"n": 1, "mean": 0.4061601614382224, "std": null, "lo": null, "hi": null, "values": [0.4061601614382224]},
    "brier": {"n": 1, "mean": 0.185728354790218, "std": null, "lo": null, "hi": null, "values": [0.185728354790218]},
    "temperature": {"n": 1, "mean": 2.506976289492856, "std": null, "lo": null, "hi": null, "values": [2.506976289492856]},
    "ece_temp": {"n": 1, "mean": 0.0490372448507657, "std": null, "lo": null, "hi": null, "values": [0.0490372448507657]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0501232998796224, "std": null, "lo": null, "hi": null, "values": [0.0501232998796224]},
    "mce_temp": {"n": 1, "mean": 0.2399333649365403, "std": null, "lo": null, "hi": null, "values": [0.2399333649365403]},
    "brier_temp": {"n": 1, "mean": 0.1607166743611673, "std": null, "lo": null, "hi": null, "values": [0.1607166743611673]}
   }},
  {"model": "deep-res-cnn", "task": "ltafdb-binary", "averaging": "binary", "runs": 1, "seeds": [42], "plot_seed": 42, "plots": ["confusion-matrix", "calibration-curve", "loss-curve"],
   "metrics": {
    "f1": {"n": 1, "mean": 0.934156441953908, "std": null, "lo": null, "hi": null, "values": [0.934156441953908]},
    "auroc": {"n": 1, "mean": 0.9767413713583726, "std": null, "lo": null, "hi": null, "values": [0.9767413713583726]},
    "accuracy": {"n": 1, "mean": 0.9353673648260646, "std": null, "lo": null, "hi": null, "values": [0.9353673648260646]},
    "mcc": {"n": 1, "mean": 0.8742458543140129, "std": null, "lo": null, "hi": null, "values": [0.8742458543140129]},
    "precision": {"n": 1, "mean": 0.9775114375926284, "std": null, "lo": null, "hi": null, "values": [0.9775114375926284]},
    "recall": {"n": 1, "mean": 0.8944839175683246, "std": null, "lo": null, "hi": null, "values": [0.8944839175683246]},
    "ece": {"n": 1, "mean": 0.0344262859244277, "std": null, "lo": null, "hi": null, "values": [0.0344262859244277]},
    "adaptive_ece": {"n": 1, "mean": 0.0361378077347467, "std": null, "lo": null, "hi": null, "values": [0.0361378077347467]},
    "mce": {"n": 1, "mean": 0.054800573888689, "std": null, "lo": null, "hi": null, "values": [0.054800573888689]},
    "brier": {"n": 1, "mean": 0.1096010725251483, "std": null, "lo": null, "hi": null, "values": [0.1096010725251483]},
    "temperature": {"n": 1, "mean": 2.3410212451142196, "std": null, "lo": null, "hi": null, "values": [2.3410212451142196]},
    "ece_temp": {"n": 1, "mean": 0.0472449099426332, "std": null, "lo": null, "hi": null, "values": [0.0472449099426332]},
    "adaptive_ece_temp": {"n": 1, "mean": 0.0474244743304304, "std": null, "lo": null, "hi": null, "values": [0.0474244743304304]},
    "mce_temp": {"n": 1, "mean": 0.2383743400714347, "std": null, "lo": null, "hi": null, "values": [0.2383743400714347]},
    "brier_temp": {"n": 1, "mean": 0.1194709976043544, "std": null, "lo": null, "hi": null, "values": [0.1194709976043544]}
   }}
 ]
}
