Optimizing single-lead ECG axis for AI-based detection of myocardial diseases

Figures

• 

  Fig. 1. Study overview and virtual lead synthesis. The left panel illustrates the conceptual framework for synthesizing ECG leads at various orientations using vector projections derived from Leads I and II. Synthetic virtual vectors were computed by combining Leads I and aVF through directional vector synthesis, with Lead aVF itself derived from Leads I and II using the standard Einthoven and Goldberger formula. The right panel displays representative 10-s waveforms of synthesized single-lead ECGs across a range of angles.

• 

  Fig. 2. Performance of AI models across synthetic lead axes in the BWH cohort. AUROC values on the test set for disease detection models using each synthesized lead, ranging from 0° to 350° in 10-degree increments. Each point represents the mean AUROC across five independently trained models; error bars indicate 95% CIs. AUROC values for A LVSD, B HCM, and C CA tested in the BWH cohort. BWH Brigham and Women’s Hospital, LVSD left ventricular systolic dysfunction, HCM hypertrophic cardiomyopathy, CA cardiac amyloidosis, CI confidence interval.

• 

  Fig. 3. Subgroup analysis of model performance across BWH patient characteristics. Model performance for detecting LVSD, HCM, and CA was evaluated across different patient characteristics, stratified by sex (male vs. female), race (White vs. non-White), and age quartiles. The age cutoffs were: 53, 66, and 75 years for LVSD; 47, 57, and 67 years for HCM; and 67, 73, and 79 years for CA. Each point represents the mean AUROC across five independently trained models; error bars indicate 95% CIs. BWH Brigham and Women’s Hospital, LVSD left ventricular systolic dysfunction, HCM hypertrophic cardiomyopathy, CA cardiac amyloidosis, CI confidence interval.

• 

  Fig. 4. Performance of the BWH integrated model for optimal-lead axis and Lead I. AUROC curves of the best integrated models in disease detection, with the optimal-lead axis of 20° for LVSD and HCM, and 210° for CA. The results of integrated models using Lead I are shown for comparison. The 95% CIs for the AUROCs were calculated using 2000 bootstrap resamples of the test set. BWH Brigham and Women’s Hospital, LVSD left ventricular systolic dysfunction, HCM hypertrophic cardiomyopathy, CA cardiac amyloidosis, CI confidence interval.