[Paper Review] AI Generalisation Gap In Comorbid Sleep Disorder Staging

Saswata Bose, Suvadeep Maiti+5 more|arXiv (Cornell University)|Mar 24, 2026

EEG and Brain-Computer Interfaces0 citations

TL;DR

The paper demonstrates that sleep-staging models trained on healthy data poorly generalize to ischemic-stroke patients with comorbid sleep disorders, using iSLEEPS and Grad-CAM explanations to reveal misaligned focus.

ABSTRACT

Accurate sleep staging is essential for diagnosing OSA and hypopnea in stroke patients. Although PSG is reliable, it is costly, labor-intensive, and manually scored. While deep learning enables automated EEG-based sleep staging in healthy subjects, our analysis shows poor generalization to clinical populations with disrupted sleep. Using Grad-CAM interpretations, we systematically demonstrate this limitation. We introduce iSLEEPS, a newly clinically annotated ischemic stroke dataset (to be publicly released), and evaluate a SE-ResNet plus bidirectional LSTM model for single-channel EEG sleep staging. As expected, cross-domain performance between healthy and diseased subjects is poor. Attention visualizations, supported by clinical expert feedback, show the model focuses on physiologically uninformative EEG regions in patient data. Statistical and computational analyses further confirm significant sleep architecture differences between healthy and ischemic stroke cohorts, highlighting the need for subject-aware or disease-specific models with clinical validation before deployment. A summary of the paper and the code is available at https://himalayansaswatabose.github.io/iSLEEPS_Explainability.github.io/

Motivation & Objective

• Highlight the generalization gap when applying healthy-subject sleep staging models to ischemic stroke patients with comorbid sleep disorders.
• Introduce iSLEEPS, a clinically annotated PSG dataset of stroke patients, to enable pathology-aware benchmarking.
• Develop and benchmark a SE-ResNet plus bidirectional LSTM model for single-channel EEG sleep staging.
• Use Grad-CAM explainability and clinician feedback to assess model attention and clinical relevance.

Proposed method

• Process raw single-channel EEG with a sliding 30 s window to predict the central epoch's sleep stage.
• Use SE-ResNet to extract spectral–temporal features and suppress noise.
• Incorporate stacked Bi-LSTM layers to model bidirectional temporal dependencies.
• Train and evaluate models on four datasets (SleepEDF-20, SleepEDF-78, SHHS, iSLEEPS) with leave-one-out cross-validation.
• Apply Grad-CAM to generate heatmaps that indicate which EEG regions drive decisions, validated by clinicians.

Fig. 1 : Deep Model Architecture, with the $n$ th epoch being input to the model, having a window size of $w$ and a stride length of $s$ .

Experimental results

Research questions

• RQ1 Do models trained on healthy sleep data generalize to ischemic stroke patients with comorbid sleep disorders?
• RQ2 How does explainability (Grad-CAM) align with clinically meaningful EEG features in patient data?
• RQ3 What is the performance gap when applying a healthy-trained model to iSLEEPS compared to training on patient data?
• RQ4 What dataset-specific differences in sleep architecture explain the generalization failure?

Key findings

• Model trained on healthy data shows substantial generalization gap when applied to iSLEEPS (patient cohort).
• On iSLEEPS, the proposed benchmarking model achieves 74.7% ACC, 67.7 MF1, and 0.64 κ, lower than healthy-cohort results but indicating some predictive capacity.
• Grad-CAM heatmaps reveal attention to physiologically uninformative EEG regions in stroke patients when trained on healthy data.
• Clinician feedback confirms that many activations correspond to diffuse slowing, artifacts, or non-sleep features rather than spindles, K-complexes, or alpha rhythms.
• Statistical analysis shows significant differences in sleep architecture between healthy and ischemic stroke cohorts, justifying subject-aware or disease-specific models.
• A binary-classifier analysis using engineered features can perfectly separate healthy vs. patient recordings, underscoring substantial domain shift.

(a) A Correctly Tagged Epoch. The green box highlights the Alpha Waves present in the epoch.