SNN-FS: A Spiking Neural Network Feature Selection Framework for Accurate Sleep Stage Classification from EEG Signals

Accurate classification of sleep stages is essential for advancing sleep research and improving the clinical diagnosis of sleep disorders. This study proposes a robust and scalable machine learning framework for automated sleep staging based on polysomnographic (PSG) recordings comprising electroencephalogram (EEG) signals. Following initial signal extraction and preprocessing, a hybrid feature extraction strategy is employed, incorporating statistical descriptors and wavelet-based analysis using the Daubechies-4 (db4) transform to capture both temporal and frequency-domain characteristics. To address class imbalance and enhance generalizability, Gaussian Noise Data Augmentation (GNDA) is applied. Furthermore, we introduce a Spiking Neural Network (SNN)-based feature selection mechanism that leverages the biological plausibility and energy efficiency of spiking neurons to identify the most salient features while preserving temporal dynamics. The dataset was partitioned using a 70-15-15 stratified split for training, validation, and testing, ensuring representative distribution across all sleep stages. The classification task is performed using four models: XGBoost, LightGBM, CatBoost, and a convolutional neural network combined with long short-term memory (CNN-LSTM). Among these, XGBoost demonstrates superior performance, achieving an overall accuracy of 94.27% and an F1-score of 0.9427 on the hold-out test set. Comparative analysis reveals that the proposed methodology with SNN-based feature selection significantly outperforms the baseline approach without feature selection, improving the F1-score from 0.9126 to 0.9427 for XGBoost. The proposed methodology effectively mitigates overfitting, enhances class separability, and achieves robust performance across all sleep stages, underscoring its potential for deployment in real-world clinical and research environments.