Interpretable detection of left ventricular hypertrophy using commercial ECG features and machine learning: a study based on the PTB-XL+ dataset.

Introduction and objectives: Left ventricular hypertrophy (LVH) predicts cardiovascular morbidity and mortality. ECG is the most accessible LVH screening tool, but conventional voltage criteria have low sensitivity. Deep learning improves detection but lacks interpretability. This study aims to evaluate the utility of PTB-XL+ pre-extracted commercial ECG features for interpretable LVH detection. Methods: We retrospectively analyzed 11,692 PTB-XL+ ECG records (2,258 LVH, 9,434 normal). Three feature sets (GE 12SL:782, UNIG:702, ECGDeli:529) and four classifiers (XGBoost, LightGBM, Random Forest, Logistic Regression) were evaluated via five-fold stratified cross-validation. SHAP analysis provided feature importance interpretations. Results: XGBoost performed best (AUC: 12SL = 0.9859, UNIG = 0.9853, ECGDeli = 0.9828). Commercial features marginally outperformed open-source ECGDeli. SHAP identified key features (V5-V6T-wave changes, precordial R-wave amplitude, V1 QRS area) and age as an important non-ECG predictor. Conclusion: Interpretable machine learning (ML) with commercial ECG features achieves excellent LVH detection with clinically meaningful attributions, bridging high performance and interpretability for clinical integration.