Age prediction from 12-lead electrocardiograms using deep learning: a comparison of four models on a contemporary, freely available dataset

Abstract

Objective. The 12-lead electrocardiogram (ECG) is routine in clinical use and deep learning approaches have been shown to have the identify features not immediately apparent to human interpreters including age and sex. Several models have been published but no direct comparisons exist. Approach. We implemented three previously published models and one unpublished model to predict age and sex from a 12-lead ECG and then compared their performance on an open-access data set. Main results. All models converged and were evaluated on the holdout set. The best preforming age prediction model had a hold-out set mean absolute error of 8.06 years. The best preforming sex prediction model had a hold-out set area under the receiver operating curve of 0.92. Significance. We compared performance of four models on an open-access dataset.