Machine learning to classify left ventricular hypertrophy using ECG feature extraction by variational autoencoder.

Background: Traditional ECG criteria for left ventricular hypertrophy (LVH) have low diagnostic yield. Machine learning (ML) can improve ECG classification.

Methods: ECG summary features (rate, intervals, axis), R-wave, S-wave and overall-QRS amplitudes, and QRS/QRST voltage-time integrals (VTIs) were extracted from 12-lead, vectorcardiographic X-Y-Z-lead, and root-mean-square (3D) representative-beat ECGs. Latent features were extracted by variational autoencoder from X-Y-Z and 3D representative-beat ECGs. Logistic regression, random forest, light gradient boosted machine (LGBM), residual network (ResNet) and multilayer perceptron network (MLP) models using ECG features and sex, and a convolutional neural network (CNN) using ECG signals, were trained to predict LVH (left ventricular mass indexed in women >95 g/m², men >115 g/m²) on 225,333 adult ECG-echocardiogram (within 45 days) pairs. AUROCs for LVH classification were obtained in a separate test set for individual ECG variables, traditional criteria and ML models.

Results: In the test set (n=25,263), AUROC for LVH classification was higher for ML models using ECG features (LGBM 0.790, MLP 0.789, ResNet 0.788) as compared to the best individual variable (VTI 0.677), the best traditional criterion (Cornell voltage-duration product 0.647) and CNN using ECG signal (0.767). Among patients without LVH who had a follow-up echocardiogram >1 (closest to 5) years later, LGBM false positives, compared to true negatives, had a 2.63 (95% CI 2.01, 3.45)-fold higher risk for developing LVH (p<0.0001).

Conclusions: ML models are superior to traditional ECG criteria to classify-and predict future-LVH. Models trained on extracted ECG features, including variational autoencoder latent variables, outperformed CNN directly trained on ECG signal.