GradG-AttnSleep: Dual-Mode GradCAM for Interpretable Sleep Staging with an Attention-Based CNN
Issued Date
2025-01-01
Resource Type
eISSN
27660419
Scopus ID
2-s2.0-105034211295
Journal Title
Proceedings of the International Conference on Information Technology and Electrical Engineering Icitee
Rights Holder(s)
SCOPUS
Bibliographic Citation
Proceedings of the International Conference on Information Technology and Electrical Engineering Icitee (2025)
Suggested Citation
Mahmood S.N., Ferdous T., Zereen A.N. GradG-AttnSleep: Dual-Mode GradCAM for Interpretable Sleep Staging with an Attention-Based CNN. Proceedings of the International Conference on Information Technology and Electrical Engineering Icitee (2025). doi:10.1109/ICITEE66631.2025.11338388 Retrieved from: https://repository.li.mahidol.ac.th/handle/123456789/115961
Title
GradG-AttnSleep: Dual-Mode GradCAM for Interpretable Sleep Staging with an Attention-Based CNN
Author(s)
Author's Affiliation
Corresponding Author(s)
Other Contributor(s)
Abstract
Deep learning models have shown superior performance in EEG-based automated sleep stage classification, closely matching clinical assessments of sleep quality and sleep disorder diagnosis. However, despite impressive results from models like AttnSleep, limited interpretability restricts their clinical adoption. This study proposes a framework that integrates a modified Gradient-weighted Class Activation Mapping (Grad-CAM) as an explainability mechanism within the AttnSleep architecture. The dual-mode Grad-CAM technique highlights critical EEG segments that influence each sleep stage decision, improving model transparency and classification accuracy. The method was evaluated on the Sleep-EDF-20 dataset using single-channel FpzCz EEG recordings. The model achieved an overall accuracy of 87.33%, with class-wise F1-scores of Wake: 93.35, N1: 43.16, N2: 92.55, N3: 93.61, and REM: 86.98. Compared to previous explainable models such as SleepXAI, the proposed method improves classification performance in most stages, except N1, which remains challenging due to its transient nature and overlap with other stages. These findings highlight the need to incorporate explainability into deep models for accurate, clinically relevant sleep stage classification.