SleepXNet: DeepSeek-Guided Hybrid Transformer Model for Clinically Interpretable EEG-Based Sleep Stage Classification

Sleep-stage classification using EEG signals is crucial for understanding brain activity during sleep; however, the black-box nature of most deep learning models limits their clinical adoption and practical use. This study presents a novel hybrid framework that enhances both classification accuracy and interpretability. The pipeline begins with Spectrum Attention-Based Bandpass Filtering (SABF), which denoises EEG signals and preserves important spectral features. A conditional GAN is used to address class imbalance and missing data by generating synthetic segments. The feature extraction process uses a dual-path strategy: (i) the Fourier-Temporal Fusion Network (FTFNet) captures time-frequency and inter-channel relationships by utilizing TKABNet and FTLNet, and (ii) parallel CNN streams extract spatially complementary features from both raw and wavelet-transformed EEG. These outputs are fused to form a unified representation that supports rich abstraction. The fused features are passed through a Transformer encoder to model long-range dependencies, a bidirectional GRU (Bi-GRU) for local temporal continuity, and a Graph Attention Network (GAT) to capture physiological transitions. The model feeds the combined output into a fully connected layer that uses softmax to predict the final sleep stage. For interpretability, the DeepSeek module integrates SHAP-based attribution, attention map visualization, and neuro-symbolic reasoning (e.g., spindle $\rightarrow$ N2). DeepSeek also generates clinician-style reports with visual and textual justifications. Evaluations on the SleepEDF-20, CAP, and MASS-SS3 datasets show that the model achieves the highest Accuracy on SleepEDF-20 (88.1%) and CAP (82.0%), while maintaining competitive Accuracy on MASS-SS3 (87.5%), demonstrating strong generalization, transparency, and readiness for clinical deployment.