Automated Sleep Stage Classification Using PSO-Optimized LSTM on CAP EEG Sequences.

Area of Science:

• Computational neuroscience
• Artificial intelligence in medicine

Background:

• Automatic sleep stage and Cyclic Alternating Pattern (CAP) subtype classification from EEG is challenging due to short event durations and class imbalance.
• Existing methods struggle with the complexity of sleep microstructure analysis.

Purpose of the Study:

• To develop a domain-specific deep learning system for accurate EEG-based sleep stage and CAP subtype classification.
• To enhance model performance and interpretability using hybrid optimization techniques.

Main Methods:

• Implementation of a Long Short-Term Memory (LSTM) network.
• Optimization using a Particle Swarm Optimization-Hyperband (PSO-Hyperband) hybrid hyperparameter tuning method.
• Application of SHAP-based interpretability techniques for feature analysis.

Main Results:

• Achieved 97% accuracy for REM and 96% accuracy for stage S0 on the CAP Sleep Database.
• Obtained ROC AUC scores exceeding 0.92 for challenging CAP subtypes (A1-A3).
• Demonstrated improved model transparency by identifying key spectral and morphological EEG features.

Conclusions:

• The proposed framework effectively handles class imbalance and improves discrimination between similar CAP subtypes.
• Hybrid optimization methods enhance the performance, generalizability, and interpretability of deep learning models for sleep microstructure analysis.