Skip to content

Latest commit

 

History

History
38 lines (32 loc) · 3.72 KB

File metadata and controls

38 lines (32 loc) · 3.72 KB

Baseline detector — per-label AUROC (validation fold)

1D ResNet, class-weighted BCE. Generated by python -m src.detection.train.

  • Macro-AUROC: 0.9141 (over 71 labels with both classes present in val)
  • Macro-F1 0.3741 · Micro-F1 0.5846 · ECE 0.9037 · n_val 2183

Calibration caveat: the class-weighted BCE (pos_weight) deliberately trades calibration for recall on rare labels, so raw probabilities run high and ECE is poor. AUROC (ranking) is the headline baseline number; probability calibration (temperature scaling) and the per-label operating thresholds are follow-up work. F1 above already uses per-label thresholds tuned on this fold.

Best 10 labels

code description train_support val_support auroc
TRIGU trigeminal pattern (unknown origin, SV or Ventricular) 16 2 0.997478
2AVB second degree AV block 12 1 0.99725
BIGU bigeminal pattern (unknown origin, SV or Ventricular) 66 8 0.996149
PSVT paroxysmal supraventricular tachycardia 19 3 0.996024
CRBBB complete right bundle branch block 432 55 0.99589
CLBBB complete left bundle branch block 428 54 0.993146
PACE normal functioning artificial pacemaker 237 29 0.991451
STACH sinus tachycardia 661 83 0.991423
PVC ventricular premature complex 915 114 0.990346
AFIB atrial fibrillation 1211 151 0.986361

Weakest 10 labels (scored)

code description train_support val_support auroc
ISCLA ischemic in lateral leads 113 14 0.837516
LVOLT low QRS voltages in the frontal and horizontal leads 145 19 0.82418
VCLVH voltage criteria (QRS) for left ventricular hypertrophy 701 87 0.802569
LAO/LAE left atrial overload/enlargement 341 43 0.790872
IVCD non-specific intraventricular conduction disturbance (block) 630 78 0.777818
ABQRS abnormal QRS 2683 322 0.752242
HVOLT high QRS voltage 49 7 0.715074
STE_ non-specific ST elevation 22 3 0.71315
TAB_ T-wave abnormality 28 4 0.686324
PMI posterior myocardial infarction 13 2 0.627464

Note (Phase 17): the ECE figure on this page came from a mis-specified metric (multi-class formulation applied to independent sigmoids) and is ~11x too high. The corrected uncalibrated value is 0.079; see docs/calibration/report.md. This page is left as the original artifact.