Artificial intelligence-based automated sleep staging using heart rate variability: Assessment of performance and clinical prospects

Background Some artificial intelligence models use heart rate variability (HRV) features to classify sleep stages. Estimation of HRV indices requires the removal of outliers and correction of ectopic beats from the electrocardiography (ECG) data. In addition, sleep epochs bear a temporal relationship to each other, while classifiers treat epochs as independent entities. To address this, we propose a bidirectional long short-term (BLSTM) architecture and a random forest classifier (RFC) for the classification of 5 sleep stages, using HRV features and sleep epoch index. Methods Polysomnography data of 645 subjects from the 'You Snooze You Win: The PhysioNet/Computing in Cardiology Challenge 2018' dataset were selected. ECG data were corrected for outliers and ectopic beats using linear interpolation. Time, frequency, and non-linear domains' HRV indices were determined and fed into the models. Results The RFC and BLSTM models obtained validation classification accuracy of 79.6 (1.6) and 74.70 (1.05), respectively, with RFC outperforming BLSTM. Further, the RFC model was validated with polysomnography data of 43 subjects from the Haaglanden Medisch Centrum dataset and obtained an accuracy of 78.9% with Cohen's kappa of 0.70 and macro F1 score of 0.789. We also demonstrated the importance of a proper pre-processing pipeline and incorporating the epoch index as a temporal feature for training the classification model. Conclusion Automated analysis of sleep data from diverse centres has potential to provide insights into population-wide sleep patterns and disease stratification, enabling targeted interventions and public health education.