Abstract Body:
Background: Federated learning (FL) enables privacy-preserving collaborative model training across healthcare institutions without sharing patient data. We evaluated disease-specific detection performance and demographic disparities in FL algorithms for ECG-based cardiovascular diagnosis across 20 cardiac conditions.
Methods: Using the PTB-XL dataset (n=21,799 ECG recordings, 18,869 patients, 20 cardiac conditions), we trained a ResNet-1D deep learning model for multi-label classification. We assessed performance across critical conditions (atrial fibrillation, myocardial infarctions, AV blocks), rhythm disorders, and conduction abnormalities. Model evaluation included disease-specific AUROC, sensitivity, and clinical priority stratification across sex and age subgroups.
Results: On 3,343 independent test samples, the model achieved mean AUROC of 0.853 across 20 conditions. Rhythm disorders showed excellent performance: sinus tachycardia (0.976), right bundle branch block (0.970), atrial flutter (0.970), and sinus bradycardia (0.953). Among myocardial infarctions, anteroseptal MI achieved 0.950 AUROC with 53.9% sensitivity, while inferior MI achieved 0.889 AUROC with 62.8% sensitivity. However, critical limitations emerged: atrial fibrillation showed only 24.7% sensitivity despite 0.881 AUROC, missing 75% of cases—a major clinical concern given stroke risk. Rare conditions suffered from class imbalance: third-degree AV block (n=4, 0% sensitivity) and second-degree AV block (n=2, 0% sensitivity). Demographic analysis revealed minimal sex disparity (0.09% accuracy difference) but age-related performance decline in elderly populations (>80 years).
Conclusion: Deep learning achieves strong disease-specific detection for common cardiovascular conditions (AUROC >0.95 for rhythm disorders), but class imbalance severely limits rare condition detection. Low atrial fibrillation sensitivity represents a critical safety gap requiring intervention strategies. Federated learning offers promise for multi-site collaborative diagnosis while preserving patient privacy, though disease-specific validation is essential.
Clinical Implications: Models excel at detecting frequent conditions but fail on rare, life-threatening diseases. Clinical deployment requires condition-specific performance thresholds, class balancing strategies (oversampling, focal loss), and mandatory rare condition review protocols.