Aims: We investigate the relationship between non-linear heart rate (HR) variability (HRV) metrics and sleep stages. Overnight HRV features serve as inputs to a deep neural network for supervised classification of sleep stages. Accurate detection of sleep stages is crucial for diagnosing and monitoring sleep disorders.
Methods: We analyzed 4158 overnight polysomnography measurements from the Sleep Heart Health Study 1 (SHHS1) dataset from National Sleep Research Resource (NSRR), aligning sleep stage annotations with RR intervals from electrocardiogram signals. Dynamic detrended fluctuation analysis (DDFA) was applied to extract time- and scale-dependent exponents alpha(t,s). Mean and standard deviation of alpha(t,s) and HR across scales were used as features. We trained a Support Vector Machine (SVM) classifier using 5-fold cross-validation on a 20% data split to serve as a benchmark for our deep neural network results.
Results: Figure 1 shows an example of subject's overnight data, with distinct patterns in alpha(t,s), HR, and sleep stages. For classification, stages N1–N4 were grouped as non-REM sleep, yielding three classes: Wake, REM, and NREM, with SVM benchmark accuracies of 50%, 42%, and 60%, respectively. These are further improved with deep learning networks.