Skip to content

Data

Source

MIT-BIH Arrhythmia Database (48 records, 360 Hz, ~110k beats), loaded with the wfdb package. Single lead MLII first.

AAMI EC57 mapping to N / V / S / F / Q

  • N: normal (N), left/right bundle branch block (L, R), atrial escape (e), nodal escape (j)
  • V: premature ventricular contraction (V), ventricular escape (E)
  • S: atrial premature (A), aberrated atrial premature (a), nodal premature (J), supraventricular premature (S)
  • F: fusion of ventricular and normal (F)
  • Q: paced (/), fusion of paced and normal (f), unclassifiable (Q)

Life-threatening focus: V class (PVCs, runs suggesting VT/VF risk).

Preprocessing (planned)

  1. Baseline wander removal (median filter, 200 ms + 600 ms windows).
  2. R-peak centered beat windows, fixed length (e.g. 216 samples ≈ 600 ms). Record-level multi-label windows (e.g. 10 s) come later if multi-label target confirmed.
  3. Z-score normalization per window.
  4. Train/test split by patient (de Chazal style: DS1/DS2), never random beats.
  5. Class imbalance handled by weighted sampling and/or pos-weighted BCE, not naive oversampling.

Scalogram (planned)

  • CWT with complex Morlet, ~32 scales covering roughly 0.5–45 Hz.
  • Log-magnitude, per-window normalized.
  • Precomputed and cached; regenerated only if scale range changes.

What is NOT decided yet

  • Single-label per beat vs true multi-label per window (see README open question).
  • Exact window length and scale count (to be fixed after first data probe).