Inside the Tricog CardioCheck algorithm.
Most single-lead algorithms do one thing: they look for rhythm. Heart rate, R-R interval variability, sometimes atrial fibrillation. That narrow scope is why single-lead ECG has historically been treated as a rhythm gadget rather than a screening instrument.
TCC takes a different position. It treats the 30-second trace as a signal dense enough to carry structural and ischaemic information — and reads that signal at three resolutions simultaneously.
Two mechanisms in one model
One model, two jobs.
- The microscope. At high resolution the model scrutinises sub-clinical waveform detail — QRS complex morphology, P-wave presence and shape, localised conduction anomalies. This is where the earliest silent indicators live, in patients who present as entirely stable.
- The sieve. At low resolution it evaluates rhythm across the whole trace and sorts the queue. Nothing waits on a specialist read: the output is a category, at the point of capture, in about ten seconds.
Three scales
- Fine-grained scale. QRS morphology, P-wave detection, conduction anomalies.
- Medium scale. Intermediate structural and ischaemic indicators — ST-segment deviation and T-wave abnormality — the band that bridges subtle waveform shifts and overt disease.
- Coarse scale. The macroscopic view: rhythm consistency and long-range trend across the full 30 seconds, catching overarching irregularity that local windows miss.
The three scales run in parallel, not in sequence. Features from every horizon are fused before classification, so ambiguity at one resolution can be resolved by context from another.
Why multi-scale
A single-resolution model forces a trade-off. Tune it for fine detail and it over-reads noise; tune it for rhythm and it goes blind to ST-segment and T-wave change. Reading all three horizons at once removes the trade-off instead of splitting the difference.
It also mirrors how a cardiologist reads a tracing: sweep for rhythm, interrogate morphology beat by beat, then step back and judge the whole. The architecture encodes that workflow rather than approximating its output.
Three parallel temporal scales — fine, medium and coarse — fused into a single classification.
AUROC 0.909 (95% CI 0.905–0.913) on held-out data.
Additive only: the model can escalate a patient's priority, never downgrade or override a clinical decision.
Trained on millions of ECGs, class-imbalance corrected.
What it detects
High risk covers time-critical findings — infarction, atrial fibrillation — where the Golden Hour is the limiting constraint. Moderate risk covers ventricular hypertrophy and ischaemic changes that aren't immediately life-threatening but require timely, focused evaluation. Low risk confirms normal sinus rhythm and benign variants.
Validation
Overall discrimination is AUROC 0.909 [95% CI 0.905–0.913] against held-out benchmarks. High-risk detection reaches 85.9% sensitivity [84.6–87.0] at 92.2% specificity [91.6–92.9]. Low-risk classification holds 90.2% sensitivity [89.2–91.2] — the figure that matters for safely routing stable patients out of critical pathways. Class imbalance was addressed explicitly in training, so rare life-threatening findings are detected with the same precision as common normal rhythms, which also mitigates algorithmic bias across demographics.
The short version
Thirty seconds of single-lead ECG, read at three resolutions at once, fused into Low, Moderate or High in about ten seconds. It escalates and never suppresses — so deploying it cannot produce worse triage than current practice, only the same or better.