Sleep Stage Classification with Non-Linear Heart Rate Variability Methods and Deep Learning

We investigate sleep stage classification using heart rate variability (HRV), comparing conventional features-mean RR, root mean square of successive differences (RMSSD), low-frequency to high-frequency power ratio (LF/HF), SD1 and SD2 of the Poincaré plot-with those derived from dynamical detrended fluctuation analysis (DDFA).We trained three classifiers: logistic regression, eXtreme Gradient Boosting, and a neural network on each feature space independently, using physicianannotated sleep stages as ground truth, labeled as wake, rapid eye movement (REM) sleep, and non-rapid eye movement (NREM) sleep.The dataset comprised 2052 subjects without a history of sleep apnea, aged 40 to 89.DDFA consistently outperformed conventional HRV, improving NREM and REM classification across all models, indicating that it provides information beyond conventional HRV features for sleep stage classification.DDFA effectively captures correlations between consecutive heartbeats and are easily visualized and interpreted, allowing physiological links to different sleep stages.This interpretability makes DDFA features particularly valuable for explainable artificial intelligence, where understanding both features and model decisions is crucial.