Obstructive sleep apnea prediction from electrocardiogram scalograms and spectrograms using convolutional neural networks

Objective. In this study, we conducted a comparative analysis of deep convolutional neural network (CNN) models in predicting obstructive sleep apnea (OSA) using electrocardiograms. Unlike other studies in the literature, this study automatically extracts time-frequency features by using CNNs instead of manual feature extraction from ECG recordings. Approach. The proposed model generates scalogram and spectrogram representations by transforming preprocessed 30 s ECG segments from time domain to the frequency domain using continuous wavelet transform and short time Fourier transform, respectively. We examined AlexNet, GoogleNet and ResNet18 models in predicting OSA events. The effect of transfer learning on success is also investigated. Based on the observed results, we proposed a new model that is found more effective in estimation. In total, 152 ECG recordings were included in the study for training and evaluation of the models. Main results. The prediction using scalograms immediately 30 s before potential OSA onsets gave the best performance with 82.30% accuracy, 83.22% sensitivity, 82.27% specificity and 82.95% positive predictive value. The prediction using spectrograms also achieved up to 80.13% accuracy and 81.99% sensitivity on prediction. Per-recording classification suggested considerable results with 91.93% accuracy for prediction of OSA events. Significance. Time-frequency deep features of scalograms and spectrograms of ECG segments prior to OSA events provided reliable information about the possible events in the future. The proposed CNN model can be used as a good indicator to accurately predict OSA events using ECG recordings.

Automated summary

This study compared different deep convolutional neural network models for predicting OSA using ECG, proposing a more effective model that automatically extracts time-frequency features and transforms them. Prediction using scalograms outperformed spectrograms, with overall high accuracy of 91.93% for per-recording classification of OSA events.