AI GENERALISATION GAP IN COMORBID SLEEP DISORDER STAGING

Abstract

Accurate sleep staging is essential for diagnosing disorders like obstructive sleep apnea (OSA) and hypopnea, both highly prevalent among stroke patients. Conventional polysomnography (PSG), though clinically reliable, is labor-intensive, costly, and reliant on manual scoring. Recent advances in deep learning enable automated staging from raw EEG of healthy participants. However, our investigations reveal that most models fail to generalize to clinical populations with disrupted sleep architecture. This paper is meant to demonstrate this phenomenon in a structured manner highlighting this issue through Grad-CAM based interpretations. To do this, we introduce iSLEEPS, a new clinically annotated ischemic stroke dataset (to be made public in near future), and evaluate a deep learning framework combining SE-ResNet feature extraction with bidirectional LSTM layers for contextual sleep stage classification using single-channel EEG. As expected, the model indicates poor cross-domain generalization between healthy and diseased patients. Comparison of the attention visualisations with expert clinical feedback reveals that this might be due to the model attending to physiologically uninformative EEG regions in patient data. Further supportive statistical and computational analyses confirm significant differences in sleep architecture between healthy and ischemic stroke subjects. These findings highlight the need for subject-aware or disease-specific sleep staging models that focus on medically relevant features, alongside clinical validation of automated tagging before integration into treatment workflows.