Electrocardiogram Abnormality Classification From Summary Statistics With Patient‐Level Validation and Process‐Focused Evaluation

Figures

• 

  FIGURE 1. Temporal data leakage in ECG abnormality classification. Valid training features (left) consist of objective ECG measurements and patient demographics available at the time of recording. The diagnosis text field (right) represents temporal leakage as it contains the cardiologist's post‐interpretation clinical conclusion, which directly reveals the classification target. Including such post‐diagnostic features in ML training creates artificially inflated performance metrics that cannot be replicated in real‐world deployment, where the model must predict abnormality from raw measurements before clinical interpretation occurs. Figure concept adapted from Toma (2025).

• 

  FIGURE 2. Patient‐level data leakage in ECG abnormality classification. Left: Proper patient‐level data splitting ensures all ECG recordings from the same patient remain in the same dataset partition (training, validation, or test). Right: Inappropriate observation‐level splitting allows multiple recordings from the same patient to appear across different partitions, leading to data leakage where the model learns patient‐specific patterns rather than generalizable ECG abnormality features. This artificially inflates performance estimates that fail to generalize to new patients in clinical deployment. Figure concept adapted from Toma (2025).

• 

  FIGURE 3. Confusion matrices for Scenario 1 (Abnormal vs. Normal, excluding ‘Borderline’) across training, validation, and test sets. Darker cells indicate correct predictions (True Positives and True Negatives), while lighter cells indicate misclassifications (False Positives and False Negatives).

• 

  FIGURE 4. Confusion matrices for Scenario 2 (Abnormal + Borderline vs Normal) across training, validation, and test sets. Darker cells indicate correct predictions (True Positives and True Negatives), while lighter cells indicate misclassifications (False Positives and False Negatives).

• 

  FIGURE 5. Confusion matrices for Scenario 3 (Abnormal vs Normal + Borderline) across training, validation, and test sets. Darker cells indicate correct predictions (True Positives and True Negatives), while lighter cells indicate misclassifications (False Positives and False Negatives).

• 

  FIGURE 6. Performance metrics evolution across dataset partitions for Scenario 1 (Abnormal vs Normal classification). The y‐axis starts at 50% to better visualize performance differences and trends.

• 

  FIGURE 7. Performance metrics evolution across dataset partitions for Scenario 2 (Abnormal + Borderline vs. Normal classification). The y‐axis starts at 50% to better visualize performance differences and trends.

• 

  FIGURE 8. Performance metrics evolution across dataset partitions for Scenario 3 (Abnormal vs Normal + Borderline classification). The y‐axis starts at 50% to better visualize performance differences and trends.

• 

  FIGURE 9. Model training and evaluation curves for Scenario 1 (Abnormal vs Normal classification). (a) Learning curves showing training and validation accuracy progression. (b) Precision‐recall curves for training, validation, and test sets.

• 

  FIGURE 10. Model training and evaluation curves for Scenario 2 (Abnormal + Borderline vs Normal classification). (a) Learning curves showing training and validation accuracy progression. (b) Precision‐recall curves for training, validation, and test sets.

• 

  FIGURE 11. Model training and evaluation curves for Scenario 3 (Abnormal vs. Normal + Borderline classification). (a) Learning curves showing training and validation accuracy progression. (b) Precision‐recall curves for training, validation, and test sets.

• 

  FIGURE 12. Reliability diagrams for test set predictions across all classification scenarios and calibration methods. Each row corresponds to one scenario (top: Scenario 1, middle: Scenario 2, bottom: Scenario 3), and each column shows a calibration method (left: Uncalibrated, center: Platt scaling, right: Isotonic regression). The diagonal dashed line represents perfect calibration. Deviation below the diagonal indicates overconfidence (predicted probabilities exceed observed frequencies), while deviation above indicates underconfidence. Expected Calibration Error (ECE) values are shown in subplot titles.