Detecting sleep stages after stroke using wearable sensors: machine learning design and challenges.

OBJECTIVE: Individuals with stroke often experience poor sleep, which negatively impacts health and quality of life. Assessing time spent in different sleep stages (such as deep or rapid eye movement sleep) is critical because these stages play distinct roles in brain health. Frequent, high-resolution sleep assessments could improve early diagnosis and treatment of sleep disorders during recovery. However, polysomnography (PSG), the gold standard, is difficult to implement in inpatient settings due to cost, complexity and patient burden. Wearable sensors paired with machine learning models offer a scalable alternative, but stroke-related physiological changes reduce model accuracy, and post-stroke PSG data are scarce. To address these gaps, we explored sleep-stage classifier models trained using different data combinations from inpatient, chronic and non-stroke populations. METHODS AND ANALYSIS: 14 sleep-stage classifiers were developed from three datasets: 8 inpatients undergoing stroke rehabilitation, 131 individuals with chronic stroke and 145 non-stroke controls. Ordinal logistic regression models classified sleep into two, three or four stages using heart rate and R-R interval features from ECG or blood oxygen saturation (SpO2/SaO2). RESULTS: Leave-one-subject-out models trained on control and/or chronic stroke ECG improved detection of inpatient sleep stages over inpatient-trained models alone (Cohen's kappa=0.31, 0.24, 0.17 for 2-, 3-, 4-stage classifiers, respectively). Adding SpO2/SaO2 benefitted performance for control and chronic stroke populations (~20% per stage) but worsened it for inpatients (~50% per stage), reflecting population-specific differences. CONCLUSION: Mixed-population training data enhance post-stroke sleep stage classification. Findings support scalable, wearables-based sleep monitoring to inform personalised interventions during stroke recovery.