APEX (Arrhythmia Pattern Explainer) is a clinical decision-support tool that reads 12-lead ECG signals, detects cardiac abnormalities across 71 diagnostic categories, generates a structured plain-English clinical explanation, and flags low-confidence predictions for manual review — designed to assist clinicians, not replace them.
- Assistive, not autonomous. APEX produces a ranked list of likely diagnoses with calibrated confidence scores and a written rationale grounded in the signal. A clinician remains in the loop for every decision. Any prediction below the configured confidence threshold is explicitly surfaced for manual review rather than presented as a conclusion.
- Explainable by construction. Every generated explanation must be traceable to specific leads, time windows, and morphological features via the grounding layer (attention / saliency). Explanations that cannot be grounded are flagged as potential hallucinations and withheld.
- Multi-label. A single ECG can carry several concurrent findings (e.g. atrial fibrillation + left bundle branch block). Detection is framed as multi-label classification over the 71 SCP-ECG statement categories in PTB-XL.
- Not a diagnostic device; no regulatory (FDA/CE) clearance is claimed or sought in v1.
- No real-time / on-device inference; batch and interactive API only.
- No treatment recommendations — detection + explanation + confidence only.
- No single-lead or wearable-strip support in v1 (12-lead resting ECG only).
Primary: cardiologists, ED physicians, and cardiology fellows reviewing 12-lead ECGs. Secondary: ML/clinical-informatics researchers evaluating explainable ECG models.
PTB-XL (PhysioNet, open access, no credentialing required): ~21,800 clinical 12-lead ECG records, 10 s each, at 100 Hz and 500 Hz, with full demographic metadata and SCP-ECG diagnostic labels. Chosen over MIMIC specifically because it is freely downloadable without a credentialing / data-use agreement, which keeps the project reproducible for any contributor.
Note on record count: the widely-cited "21,837" figure is from the original v1.0.1. This project pins v1.0.3, which contains 21,799 records (a small number were removed for data quality across releases). The label space is 71 SCP-ECG statements (44 diagnostic + 19 form + 12 rhythm, with category overlaps) — verified against the downloaded
scp_statements.csv.
- Confidence thresholding with an explicit "needs manual review" state.
- Hallucination / consistency checking between the detector's labels and the generated text (the text may never assert a finding the detector did not surface).
- All outputs carry a visible "decision support — verify before acting" disclaimer.