Comparison of historical criterion and artificial intelligence in patients with left ventricular hypertrophy

Abstract

Abstract Backgrounds Diagnostic criterion of left ventricular hypertrophy (LVH) on 12-leads electrocardiogram (ECG) were historically established. However, there was no study to evaluate the criterion comparing with machine learning methods. Purpose To verify the historical criterion of LVH on ECG, and to compare them with diagnosis by machine learning. Methods First, consecutive 60 patients with LVH were recruited, and one to one matching with age and sex to patients with normal cardiac function was performed. Finally, 120 patients (69.6 ± 12.6years, 38men per group) were enrolled. LVH was defined as at least one LV wall (septum, posterior wall, apex) showed thickness over 15mm on ultrasound echocardiography. No sinus rhythm, and wide QRS cases were excluded. The accuracy of historical criterion and ECG predictors were calculated by an assessment of whether a predictor was higher/lower than the cut-off value of receivor operating characteristics curve analysis (ROC). Ten machine learning methods were built using the significant predictors of logistic regression with 10-times cross validation. Results By logistic regression analysis, 77 significant predictors were extracted, and 22 predictors showed their area under ROC (AUROC) over 0.700. Among historical criterion, Cornell voltage showed high accuracy (0.783) and AUROC (0.808). Conversely, among AI methods, light gradient boosting machine demonstrated higher accuracy (0.843) and random forest method higher AUROC (0.882). Shapley additive explanation method (SHAP) demonstrated that V2/V2 S-wave amplitude and I/V5 T-wave amplitude played essential roles to build the AI models. Conclusions AI diagnosis on ECG for LVH showed powerful diagnostic performance comparing historical criterion.