ECG Research Explorer
Inspect individual beats, compare models, and understand where they fail.
Heartbeat waveform
Time relative to reference beat (seconds) · Amplitude (mV)
Read the evidence
N: normal-group beats · V: ventricular ectopic · S: supraventricular ectopic · F: fusion · Q: paced or unclassifiable.
A high probability is a model score, not calibrated clinical confidence. The measured baseline's high accuracy hides poor performance on several minority classes. The curated examples here cannot estimate overall performance.
Trained-model inference is available in the repository's local FastAPI application. This public static demo does not run those models. The RR ablation needs earlier beat timing and is available in measured comparisons only. Beat locations are supplied; R-peak detection is not evaluated.
What the models actually achieved
20,934 beats from nine held-out patients. Epoch selection uses training-subject validation. These aggregate results are separate from the synthetic waveform examples above.
| Model | Macro-F1 | 95% patient interval | Accuracy |
|---|---|---|---|
| RF morphology | 35.66% | 29.97%–38.13% | 94.35% |
| RF morphology + RR | 37.28% | 32.00%–42.74% | 93.78% |
| CNN | 30.52% | 22.45%–35.90% | 77.20% |
The RR improvement is modest and uncertainty intervals overlap. Minority classes remain weak. High accuracy does not establish clinical usefulness.
Rows: reference labels. Columns: model predictions. Counts come from the measured holdout.
Frozen-model external testing covered all 75 INCART records (175,867 accepted beats); seven out-of-range annotations were explicitly excluded and audited.
Read the measured reports, figures and limitations ↗