Machine Learning Determines the Impact of Electrocardiography Lead Reduction on Diagnostic Accuracy

Abstract 2305066, presented at Western Atrial Fibrillation Symposium 2026

Wearable monitors increasingly rely on reduced-lead electrocardiography (ECG) systems, yet the diagnostic vulnerability of individual ECG phenotypes to lead loss remains not well defined. Determining which conditions experience diagnostic degradation when transitioning from 12-lead to 6-, 4-, 2-, and 1-lead configurations is essential for guiding the safe and effective use of reduced-lead technologies.Raw 10-second ECG segments (N = 45,152; 5000×12 samples each) from the PhysioNet 12-Lead ECG Database with 102 phenotypes were filtered to 27 diagnoses or phenotypes having ≥10 positive samples. For each diagnosis or phenotype, a random forest classifier was trained using an 80/20 train–test split across five lead configurations: 12-lead, 6-lead (I, II, V1-V4), 4-lead (I, II, V1, V2), 2-lead (II, V2), and 1-lead (II). Performance was evaluated using the area under the curve (AUC) of the receiver operating characteristic. Diagnostic vulnerability was calculated as ΔAUC = AUC (12-lead) – AUC(X-lead), where values near zero indicate high diagnostic robustness.The average diagnostic performance decreased with lead reduction (mean AUC: 12-lead 0.74, 6-lead 0.72, 4-lead 0.7, 2-lead 0.70, 1-lead 0.67). Reductions from 12 to 6, 4, and 2 leads were not statistically significant (p > 0.05), whereas performance dropped significantly with 1-lead ECGs (W = 85, p = 0.01). Diagnostic vulnerability to reduced leads varied by diagnosis or phenotype. Rhythm disorders such as atrial fibrillation (AUC: 0.79) and atrial flutter (0.65) showed minimal degradation (ΔAUC ≤0.03). Most conduction disorders also remained stable to lead reduction, such as left bundle branch block (AUC: 0.944, ΔAUC ≤0.009). Morphology disorders showed significant declines, such as left ventricular hypertrophy (AUC: 0.851, ΔAUC: 0.252 from 12 to 1 lead) and abnormal Q wave (AUC: 0.764, ΔAUC: 0.261 from 12 to 2 leads). Overall, conditions dependent on spatial QRS morphology were most sensitive to lead loss, whereas rate, rhythm, and conduction-based conditions remained relatively stable.Rate, rhythm, and conduction-based conditions remain robust with fewer leads, whereas morphology-dependent disorders degrade significantly. These findings inform which diagnoses are appropriate for wearable ECGs, and which still require full 12-lead acquisition for safe clinical interpretation.