In diesem Semester wird die im WiSe 2022/23 erfolgreich gestartete Vortragsreihe Medical Information Sciences fortgesetzt. Renommierte Wissenschaftlerinnen und Wissenschaftler unterschiedlicher Fachdisziplinen und Forschungsstandorte geben jeden Donnerstag ab 16:00 Uhr Einblicke in aktuelle Fragestellungen und Anwendungsgebiete des breiten Forschungsfeldes Medical Information Sciences.

Sleep is a fundamental physiological process, and its objective assessment is central to diagnosing a wide range of disorders, from insomnia and sleep apnoea to neurological and cardiovascular conditions. The clinical gold standard, polysomnography (PSG), requires an overnight stay in a sleep laboratory with more than a dozen electrodes recording brain activity (EEG), eye movements (EOG), muscle tone (EMG), respiration, and cardiac activity. From these signals, trained experts manually classify each 30-second epoch of the night into sleep stages — Wake, Light Sleep, Deep Sleep, and REM. While accurate, this procedure is expensive, labour-intensive, and poorly suited to long-term or large-scale monitoring.

A growing body of research therefore explores whether sleep staging can be performed from a much simpler signal: the electrocardiogram (ECG). The heart is modulated by the autonomic nervous system, which in turn reflects the state of the sleeping brain — heart rate, heart-rate variability, and respiration-induced cardiac patterns all change systematically across sleep stages. Combined with modern deep learning, this allows surprisingly accurate sleep stage classification from a single ECG channel, opening the door to wearable, low-burden, long-term sleep monitoring in both clinical and everyday settings.

This talk presents recent work on a frequently overlooked aspect of such models: the influence of sex in the training data on model performance. While well-designed ECG-based sleep staging models typically report little to no bias between male and female subjects at inference time, the effect of the sex distribution within the training set itself has rarely been studied systematically. Using a U-Net architecture retrained on data from nearly 5800 participants of the Sleep Heart Health Study, we compare models trained on female-only, male-only, and mixed-sex cohorts and evaluate their performance across sex-stratified test sets. The results reveal asymmetric generalisation behaviour that cannot be fully explained by differences in sleep stage distribution or sleep-disorder prevalence between men and women, pointing to deeper, as-yet-unidentified physiological or signal-level differences. At the same time, a model trained on a balanced mixed-sex cohort performs equally well on both sexes, suggesting that — for the practical classification task — sex-specific models may not be necessary.