Anti-correlation time-frequency synergy and dual-granularity fusion for sleep stage classification

Automated sleep stage classification is crucial for diagnosing sleep disorders and evaluating sleep quality. Single-channel EEG sleep staging using central electrodes has garnered significant interest due to its ability to detect a higher number of sleep spindles and its cost-effectiveness. Nevertheless, several challenges remain, such as the neglect of interdependencies among features, reduced model sensitivity to the dispersion across features, potential information leakage during the learning of transition rules, and class imbalance.

To address these challenges, this study specifically designs an automatic sleep staging network based on single-channel EEG. The network incorporates a multi-branch feature extraction module, which captures negatively correlated frequency features that exhibit significant variations during sleep transitions and promotes synergistic complementarity. Furthermore, a dual-granularity residual attention mechanism is employed to perform adaptive fusion along both temporal and channel dimensions based on statistical measures such as the mean and standard deviation. The causal transition-aware network ensures temporal consistency while facilitating effective learning of transition rules. The focal loss function is utilized to counteract class imbalance.

Experiments on SleepEDF-20, SleepEDF-78 and SHHS datasets show that our method achieves the best performance with overall accuracy of 87.0%, 83.4% and 84.5%, respectively. For N1 stage classification which is more challenging, it reaches an accuracy of 57.5%, 52.5% and 49.3% on three datasets, respectively.

The experimental results demonstrate the effectiveness of the proposed approach and its certain advantages over existing methods.