Clinical teams
For cardiologists and ICU technicians who need explainability, physical quality gates, and comparable historical outcomes.
- Lead-level saliency
- Detached-lead rejection
- Review-priority triage
Self-supervised ECG representations for clinical triage.
Why annotate thousands of waveforms? CardioRep trains on raw, unlabelled 12-lead ECGs using masked prediction—reducing annotation requirements by up to 99% while enabling explainability and similar-case retrieval.
0.7936 AUROC at 1% labels17,441 indexed cases12-lead attribution
Expert labels do not scale.
Traditional medical AI depends on scarce cardiologist hours, costly annotations, and labels shaped by inter-observer variance.
→
Learn from the signal itself.
Masking intervals and leads forces CardioRep to model rhythm, transition, and vector-cardiographic lead projections from raw waveforms.
From waveform to review-ready evidence.
A simulated walkthrough of the operational triage flow. Upload, quality gate, risk score, attribution, then grounded retrieval—no single opaque verdict.
⌁
Drop a 12-lead ECG waveform
WFDB .hea/.mat pair or a standard numerical array
Verified benchmarks, not a demo claim.
Four views of whether the representation remains useful when labels, signal quality, datasets, and demographics change.
1% LABELS · 174 RECORDS0.7936CardioRep AUROC+9.42 pts absolute vs scratch
79.36
84.1
86.8
88.6
89.3
90.4
Research boundaryRetrospective benchmark results do not establish prospective safety, calibration across hospitals, or regulatory readiness. CardioRep is a research prototype—not a medical device.
Clinical evidence. Research leverage. Product economics.
Medical AI researchers
For teams testing label efficiency, cross-dataset transfer, and SIGReg against dimensional collapse.
- Frozen-encoder transfer
- Effective-rank monitoring
- Reusable representations
Digital health builders
For founders and investors focused on cold-start data, annotation economics, and edge-wearable deployment.
- Up to 99% fewer labels
- Lower launch cost
- Compact downstream heads