Md Nahidul Islam - CoC Seminar and Panel Session

Abstract: Automatic sleep stage classification is crucial for accurate sleep assessment and the diagnosis of sleep disorders. However, current sleep stage scoring still relies heavily on visual inspection by experts, which is time-consuming, subjective, and suffers from inter-rater variability. Although deep learning (DL) methods have shown significant progress in sleep stage classification, several challenges remain: (1) most models use single-modality signals or uniform multimodal pipelines, overlooking the distinct physiological characteristics of each signal and limiting effective use of complementary information; (2) high-performing models are often computationally expensive and parameter-heavy; (3) class imbalance leads to poor recognition of minority stages; and (4) limited interpretability reduces clinical reliability. To address these issues, this research aims to develop lightweight, accurate, and interpretable DL¿based frameworks for automatic sleep stage classification using multimodal polysomnography (PSG) signals. The proposed frameworks will design modality-specific expert modules employing adaptive multi-scale convolutional operations to extract informative and complementary features from EEG, EOG, and EMG signals. Then, the extracted features will be integrated through an attention-based adaptive fusion mechanism to learn the relative importance of each modality for different sleep stages. To capture inter-epoch dependencies, a temporal context-modeling component based on recurrent and transformer architectures will be developed. To improve recognition of minority stages, data augmentation and class-balancing techniques, including amplitude scaling, time shifting, and GAN-based synthesis, will be explored. Finally, explainable AI methods such as SHAP, LIME, and Grad-CAM, together with attention visualization and gradient-based attribution, will be integrated to explain model decisions and highlight physiological patterns across sleep stages. This research is expected to achieve efficient, robust, and interpretable sleep staging with improved balance across sleep stages and enhanced clinical trust.

Join via https://unisq.zoom.us/j/85622843236?pwd=vlAWdnzJaNbxO6kLeyQvlk8FjmvN82.1&from=addon