SIGNAL INGESTION
Drop ECG Trace (.csv)
Input: Single CSV row · 187 time-sampled ECG values (360 Hz, MIT-BIH format).
Per-sample Z-score normalisation applied at inference — no stored training statistics needed.
Dataset: MIT-BIH Arrhythmia DB · 109,446 beats · 5 classes (Normal 83%, Fusion 0.7%).
history SESSION LOGS
/// No sessions yet ///
Signal Waveform Replay
/// Awaiting Signal Input ///
Inference Result
LATENCY: —READY
Upload an ECG signal to begin.
/// Awaiting inference ///
Explainability — Grad-CAM
/// Compute Heatmap to View ///
DIAGNOSTIC ARCHIVE
Dataset: MIT-BIH Arrhythmia Database (Moody & Mark, 2001) · 48 half-hour Holter recordings · 360 Hz.
5-class annotation: N (Normal), S (Supraventricular), V (Ventricular), F (Fusion), Q (Unknown/Paced).
Waveforms below are representative samples extracted during EDA (notebook 01_eda.ipynb) and saved to
data/processed/reference_waveforms.npz.
SMOTE oversampling applied on training split only to address severe class imbalance.
Normal Sinus Rhythm
Regular P-QRS-T · narrow QRS · upright T-wave.
Supraventricular Ectopic
Narrow QRS · aberrant P-wave morphology · 3% of dataset.
Ventricular Ectopic
Wide bizarre QRS · no preceding P-wave · 6% of dataset.
Fusion Beat
Hybrid N+V morphology · rarest class at 0.7% · hardest to classify.
Unknown / Paced
Paced beats or unclassifiable morphology · 7.3% of dataset.
Waveforms populated from data/processed/reference_waveforms.npz.
Run notebooks/01_eda.ipynb to generate this file from the raw MIT-BIH CSV.
Charts remain blank until that file exists.