# Deep learning to predict left ventricular hypertrophy from the electrocardiogram

## TL;DR

This study uses deep learning to accurately predict left ventricular hypertrophy from ECG data, showing strong performance in a large dataset and moderate generalizability to another population.

## Contribution

A novel deep learning model for predicting LVH from ECG data with high accuracy and comparison to prior methods and clinical benchmarks.

## Key findings

- The deep learning model achieved an AUROC of 0.97 in predicting LVH from ECG data in UK Biobank.
- The model outperformed previous supervised algorithms and clinical ECG benchmarks.
- The model showed modest generalizability to the SHIP cohort with an AUROC of 0.78.

## Abstract

Left ventricular hypertrophy (LVH) is a strong predictor of cardiovascular disease. We previously compared supervised machine learning techniques to classify cardiac magnetic resonance (CMR)-derived LVH using electrocardiogram (ECG) and clinical variables in 37 534 UK Biobank participants, obtaining an area under the receiving operating curve (AUROC) of 0.85, but with limited specificity and requiring external validation. In this study, we develop a deep learning (DL) model to improve classification with external evaluation in the Study of Health in Pomerania (SHIP).

We analysed 12-lead ECGs of 48 835 participants from the UK Biobank imaging study. The dataset was split into a training set (70%), validation set (15%), and test set (15%) for performance evaluation. The model architecture was a fully convolutional network, for which the input was the participants’ median ECG and clinical variables and the predicted indexed left ventricular mass (iLVM) as the output. A subsequent logistic regression model was used to recalibrate iLVM predictions. In UK Biobank, 717 (1.5%) participants had CMR-derived LVH and the AUROC for the DL model was 0.97. The ECG components most predictive of LVH were the QRS complex and ventricular rate. The DL model outperformed our supervised algorithms, previous DL modelling efforts and clinical ECG benchmarks. There was modest generalizability of the DL model to 1423 participants in SHIP (AUROC 0.78), with differences in clinical profile, ECG acquisition, and CMR labelling as important factors.

Our findings support the feasibility of scalable DL-based screening tools for the prediction of LVH from the ECG, whilst highlighting the need for model development using larger datasets with greater diversity to ensure generalizability.

Graphical Abstract