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

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Traditional electrocardiogram (ECG) criteria for diagnosing left ventricular hypertrophy (LVH) exhibit limited diagnostic accuracy.
• Machine learning (ML) offers a promising approach to enhance the classification capabilities of ECG data.

Purpose of the Study:

• To evaluate the efficacy of various ML models in classifying LVH using ECG features.
• To compare the performance of ML models against traditional ECG criteria and direct ECG signal analysis.
• To assess the predictive value of ML models for future LVH development.

Main Methods:

• Extracted ECG features (summary, amplitudes, voltage-time integrals) from 12-lead, X-Y-Z, and 3D ECGs.
• Utilized variational autoencoders for latent feature extraction from X-Y-Z and 3D ECGs.
• Trained and compared logistic regression, random forest, LGBM, ResNet, MLP, and CNN models on a large dataset of ECG-echocardiogram pairs.

Main Results:

• ML models utilizing extracted ECG features demonstrated superior LVH classification performance (AUROC up to 0.790) compared to traditional criteria (AUROC 0.647) and CNNs (AUROC 0.767).
• The Light Gradient Boosted Machine (LGBM) model showed high accuracy in classifying LVH.
• False positives from the LGBM model were associated with a significantly higher risk (2.63-fold) of developing future LVH.

Conclusions:

• ML models significantly surpass traditional ECG criteria for both classifying and predicting future LVH.
• Models trained on extracted ECG features, including latent variables, outperformed direct ECG signal analysis (CNN).
• ML-based ECG analysis holds substantial potential for improved cardiovascular risk assessment.