Reducing Data Dimensionality of Ecg Signal Using Deep Learning

Background Deep Learning (DL) models can predict various health conditions solely from ECG. ECG signal is large (60,000 data points) which leads to overfitting when training machine learning (ML) models with limited number of ECGs. We propose a DL cyclical β-variational autoencoder, β-VAE, to encode the representative beat ECG to 30 variables which can then be used as a surrogate for the original signal for ML. Methods We trained two β-VAEs to encode 0.75-sec representative beat ECG, vectorcardiographic XYZ-leads and root-mean-squared (RMS) signal, into 30 variables each, and reconstruct the signal back from these variables. We trained light gradient boosted machine models to predict standard ECG measurements using just the encoded variables. Results The weighted mean square error of the signal reconstructed exclusively from the encoded variables was 0.3 and 0.1 µV 2 for the XYZ and RMS respectively. The figure shows examples of the reconstructed ECG signal compared to the original. The encoded variables were able to predict QRS duration (mean absolute error 7.1 ± 12.1 ms, R 2 0.76), QT interval (18.6 ± 29.4 ms, R 2 0.65), and RMS R wave amplitude (77.8 ± 142.4 µV, R 2 0.88). Conclusion DL can encode ECG signal into a few latent variables to reduce data dimensionality. These encoded variables retain the morphological characteristics of the ECG (e.g. can predict ECG measurements within margin of measurement error) and could be used for training ML models from small datasets without overfitting.