An ECG model can explain a risk score without proving the explanation is causal

An electrocardiogram contains more information than the labels routinely read from it. CLAIRE uses 636 structured ECG features together with age and gender to predict major adverse cardiovascular events and mortality, then uses an open-source reasoning model to rank informative features and generate physiologically plausible pathways linking abnormalities to outcomes.[1]

On 60,000 adult ECGs, the full-feature model reported accuracy/AUROC of 0.97/0.98 for the composite MACE outcome and 0.86/0.86 for mortality. Board-certified physicians reviewed generated explanations for physiological plausibility. Those are strong internal results. They do not establish that the narrative explanation is the true causal path, that the score is calibrated for another hospital or that acting on it improves patient outcomes.[1]

An AUROC of 0.98 describes discrimination: how well a score ranks cases with and without the study endpoint. It does not mean that 98% of patients receive a correct individual prognosis. Accuracy, false-alarm burden and the meaning of a stated risk depend on event frequency, prediction horizon and the chosen threshold. The abstract’s headline metrics do not supply all of those clinical operating details.[1]

Prediction and explanation are different models of evidence

The predictive stage asks whether patterns in the ECG features separate people who later experience an endpoint from those who do not. The explanatory stage asks which features matter most and how they could fit a physiologically sensible story. A model can predict accurately using correlations that are not causal. A story can be medically plausible while omitting another pathway or translating a proxy into a confident mechanism.

This is not a reason to remove explanations. It is a reason to label what they are. The explanation can help a clinician inspect whether the model’s attention falls on recognisable physiology, detect nonsensical associations and formulate hypotheses. It should not be presented as proof that changing the highlighted feature would change the patient’s outcome.

A composite endpoint can conceal different tasks

MACE commonly combines several adverse outcomes. A high score for the composite does not mean the model predicts each component equally well or that each requires the same intervention. Mortality is reported separately, but deployment would need endpoint-specific definitions, horizons, calibration and action thresholds.

The feature-reduced variants are practically interesting because they retained good-to-excellent discrimination with less computational burden. A smaller feature set can also make review easier. Yet selection based on one dataset may remove signals needed in another population; reduction should be tested externally rather than treated as a universal simplification.[1]

The explanation should link back to measurable evidence

1. Show the ECG features and values that materially influenced the risk score.
2. Separate measured abnormalities from the model’s proposed physiological interpretation.
3. Report the prediction horizon, endpoint definition and calibrated risk threshold.
4. Compare the narrative with simpler attribution and feature-ablation tests.
5. Let clinicians mark an explanation as useful, misleading, incomplete or physiologically implausible.
6. Retain the original ECG and structured feature record so the explanation can be audited later.

CLAIRE is a useful advance because it does not stop at a black-box probability. It exposes candidate features and an interpretable account that clinicians can challenge. The next standard should preserve that openness while raising the evidence bar: external validation, calibration, fidelity tests and prospective decision impact. Explainability becomes clinically meaningful when it makes the prediction easier to interrogate—not when fluent physiology persuades the reader to stop asking how the model knows.

Report an error or suggest a correction