Document resource
Introduction Deep learning Artificial Intelligence models demonstrate high accuracy in electrocardiogram interpretation, but clinical adoption is limited by poor interpretability and misalignment with human expert reasoning. Eye-tracking data provides insight into how clinicians visually analyse electrocardiograms and may guide models toward clinically meaningful features. We investigated whether integrating human expert eye-tracking data into a convolutional neural network alters model attention and diagnostic performance across rhythm, ischaemia, and structural electrocardiogram phenotypes.Methods A publicly available eye-tracking dataset was used in which consultant cardiologists, fellows, and residents interpreted the same set of ten standard 12-lead electrocardiogram images. Clinicians reviewed each electrocardiogram as in routine practice while an eye-tracking system recorded fixation location and duration across predefined lead regions, producing numerical data describing visual attention to each lead.As raw voltage signals were unavailable, electrocardiogram images were converted into approximate numerical waveforms. Each image was segmented into twelve lead strips using predefined bounding boxes. For each lead, vertical pixel intensities were averaged across time, resized to 1000 time points, and normalised, resulting in a 12 × 1000 time-series representation per electrocardiogram.Clinician gaze data were aggregated by summing fixation time per lead and averaging across clinicians to derive lead-level attention profiles. Two gaze profiles were created: one using consultant data alone and one combining consultant, fellow, and resident data.An open-source convolutional neural network pre-trained on the PTB-XL electrocardiogram dataset served as the base model. Internal feature maps preserving lead and temporal structure were used to derive model attention maps. Two gaze-guided models were fine-tuned using only the ten eye-tracking electrocardiograms by optimising diagnostic performance while encouraging alignment between model and clinician attention. Only upper network layers were updated to preserve previously learned electrocardiogram features.Evaluation was conducted in two stages: internal representation analysis on PTB-XL using linear probing for myocardial infarction versus normal and atrial fibrillation versus sinus rhythm, followed by external validation on the publicly available Lobachevsky University Electrocardiography Database across rhythm, ischaemia, conduction, hypertrophy, and pacing phenotypes.Results On PTB-XL, gaze fine-tuning reduced classification accuracy compared with the base model for myocardial infarction and atrial fibrillation but produced consistent shifts in attention toward clinically relevant leads.On the Lobachevsky University Electrocardiography Database, gaze-guided models demonstrated task-dependent effects. For ischaemia versus normal electrocardiograms, the combined gaze model outperformed the base model (accuracy 0.906 vs 0.875), despite being fine-tuned on only 10 gaze-labelled electrocardiograms. For conduction abnormalities and paced rhythm, gaze-guided models achieved comparable or improved area under the receiver operating characteristic curve. Performance for atrial fibrillation was more variable.Conclusion Incorporating clinician eye-tracking data reshapes model attention toward clinically meaningful electrocardiogram features. While performance gains were task-specific, gaze-guided fine-tuning improved accuracy for selected conditions such as ischaemia using minimal training data. This supports a hybrid human–artificial intelligence approach that prioritises interpretability alongside diagnostic performance.Abstract 111 Figure 1Comparison of expert human gaze, base model attention, and gaze-fine-tuned model attention for an atrial fibrillation electrocardiogramAbstract 111 Table 1Diagnostic performance of the base model and gaze-guided models across multiple electrocardiogram phenotypes on the Lobachevsky University Electrocardiography DatabaseTaskBase model AccuracyConsultant-only gaze AccuracyAll-gaze model AccuracyAtrial fibrillation vs sinus rhythm0.8110.8680.849Ischaemia vs normal0.7000.9000.800Conduction abnormality vs normal0.7920.8490.830Left ventricular hypertrophy vs normal0.7450.7450.745Paced vs non-paced rhythm0.8670.8670.900Abstract 111 Figure 2Internal attention maps for an ischaemia-positive electrocardiogram from LUDB, comparing the base model with gaze-guided models fine-tuned using consultant gaze alone and combined consultant, fellow, and resident gaze