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

BACKGROUND

Traditional electrocardiographic (ECG) criteria for left ventricular hypertrophy (LVH) have modest diagnostic yield.

OBJECTIVE

This study aimed to develop and validate machine learning (ML) models for LVH diagnosis from ECG.

METHODS

ECG calculations (rate, intervals, and axis); R-wave, S-wave, and overall-QRS amplitudes; and QRS voltage-time integrals were obtained from 12-lead, vectorcardiographic X-Y-Z-lead, and 3-dimensional (root-sum-square) ECGs. Deep learning-enabled latent embeddings (30 per ECG) were extracted using a variational autoencoder (pretrained on unselected 1.18 million ECGs) from representative-beat signals. Logistic regression, random forest, light gradient boosted machine (LGBM), residual neural network and multilayered perceptron network models using ECG features (calculations and embeddings) and sex, and a convolutional neural network (CNN) using ECG signals alone were trained to predict LVH (left ventricular mass index, women >95 g/m2; men >115 g/m2) on 482,734 ECG-echocardiogram pairs (±45 days). Area under the receiver operating characteristic curves were reported from a holdout testing set.

RESULTS

In the testing set (n = 54,984), the area under the receiver operating characteristic curve for LVH classification was higher for ML models using ECG features (LGBM 0.794; multilayered perceptron 0.793; residual neural network 0.795) than the best individual ECG variable (Z-axis QRS voltage-time integral 0.707), the best traditional criterion (Cornell voltage-duration product 0.716), and the CNN using ECG signals (0.788). Among patients without LVH who had a follow-up echocardiogram >1 year later, LGBM false positives, compared with true negatives, had a 3.07-fold higher odds of developing future LVH (95% confidence interval 2.44-3.86; P < .0001).

CONCLUSION

ML models are superior to traditional ECG criteria for classifying LVH. Models trained on extracted ECG features, including deep-learning latent space representations, can outperform CNN models trained on ECG signals.