Teaching an ECG Model Where to Look

An electrocardiogram records the heart’s electrical activity over time, across several leads. For a machine-learning system, recognizing a diagnostic pattern is only part of the challenge: clinicians may also want to know which section of the tracing supports the prediction. This is especially difficult to train for because large ECG collections often provide diagnoses for whole recordings, not expert markings of the exact waveform regions that matter.

The paper introduces IECG, a model designed to predict diagnoses and highlight relevant parts of ECG signals together. Its central idea is to use existing clinical knowledge to make approximate training labels for those regions, rather than asking specialists to annotate every recording by hand.

Turning clinical criteria into training hints

The system starts with 12-lead ECG recordings. A signal-processing and segmentation pipeline identifies waveform components such as P waves, QRS complexes and T waves. It then applies clinical criteria—for example, rules involving interval duration or ST-segment changes—to flag candidate regions. A retrieval component connects diagnosis terms with criteria drawn from the Life in the Fastlane ECG reference. The resulting heatmaps are pseudo-labels: useful but imperfect hints, not expert annotations for each individual recording.

During training, IECG is encouraged to direct its attention toward those hinted regions. The model also learns to align ECG representations with text reports and to generate report descriptions. The authors use a loss based on Kullback–Leibler divergence to encourage attention distributions to resemble the pseudo-labels. Put simply, the model is trained not only to make a prediction, but also to look where clinical rules suggest relevant evidence may be found.

The researchers pretrained the system on MIMIC-IV-ECG, a large collection of recordings. For localization testing, they used ECGAD, a benchmark drawn from PTB-XL with expert heatmap annotations. Although the benchmark contains 400 test recordings, only 190 abnormal recordings had localization annotations. Classification was evaluated on ECGAD and on the larger PTB-XL dataset. The comparisons included other ECG models and methods designed to produce explanations or localized predictions.

Better localization in the reported tests

The paper reports that IECG’s rollout-attention variant achieved a localization AUROC of 75.55% on ECGAD, compared with 71.62% for the ECGAD model, which was built specifically for localization. AUROC measures how well a score ranks annotated abnormal points above other points; it does not mean that 75.55% of all highlighted regions were correct. The authors also report that IECG’s localization results exceeded those of the comparison methods in their evaluation.

Classification results were competitive, rather than uniformly best across every measure and dataset. On PTB-XL, the final-layer variant achieved an FMax of 42.43% and a classification AUROC of 92.21%, the highest values reported in that comparison. On ECGAD, a different IECG variant had the highest FMax, while TolerantECG had the highest classification AUROC. This distinction matters: performance depends on the metric, model variant and test set. The findings do not support a blanket claim that IECG is the best ECG classifier.

An ablation—an experiment that removes or changes one component—suggested that attention guidance was important for localization. Without it, localization AUROC fell to 61.33%, while classification remained strong. The results illustrate the intended trade-off: clinically guided attention can make regions easier to identify without necessarily improving every diagnostic score.

What the heatmaps can—and cannot—show

A highlighted region is not a complete account of a model’s reasoning. IECG produces a global attention map that can combine evidence for several diagnoses; the authors say it does not yet reliably assign each highlighted segment to a particular diagnosis in recordings with multiple conditions. Attention may help inspection, but it should not be treated as proof that a prediction is medically correct or that the highlighted waveform caused it.

The approach also inherits weaknesses from its rule-based ingredients. The authors note that the criteria database may not cover rare conditions or patient-specific factors, and that its retrieval component has limited coverage. Errors in waveform segmentation can carry forward into the generated hints. In addition, the expert localization labels used for testing may omit clinically relevant features. In the paper’s examples, some annotations missed prolonged intervals that IECG highlighted, making the model’s measured score potentially less favorable than a fuller clinical assessment would suggest. That possibility does not independently establish that the model’s extra highlights are correct.

The strongest conclusion is therefore about a method, not a clinical tool: clinical rules can provide useful weak supervision for training a model to localize ECG abnormalities. The paper reports results on established datasets, but the evidence presented here does not establish performance in routine care, across all patient populations, or under changing clinical conditions. The source is an arXiv preprint; a preprint is a publicly shared research manuscript and may not have completed peer review. Independent replication and prospective clinical evaluation would be needed before drawing conclusions about real-world use.