SleepTNet: Automatic Sleep Stage Classification With Transition Model Using Multi-Channel EEG

Sleep plays an important role in physical health, cognition, and emotional well-being. Since sleep disorders can affect all these aspects, accurately detecting sleep stages is key to proper diagnosis and monitoring. Automatic sleep stage classification is essential, as manual annotation is time-consuming and inconsistent. Most existing automatic sleep stage classification models continue to exhibit performance gaps in detecting transition stages. To address this issue, SleepTNet is proposed as a deep learning framework for automatic sleep stage classification using EEG signals. The model is composed of two modules: a representative feature extraction module and a sequential classification module. The model adopts enzyme-inspired concept and separating training strategies, with its core built around transition models designed to detect transition epochs. The proposed model contained approximately 7.7 million trainable parameters, balancing performance and model efficiency. The model was evaluated on the Massachusetts General Hospital (MGH) dataset using EEG signals and obtained an overall accuracy of 81.39%, macro-F1 score of 79.52%, and Cohen’s kappa of 0.75, outperformed several state-of-the-art results. Additionally, the model evaluated transition and non-transition epochs separately, achieved a transition epoch accuracy of 62.06% and non-transition accuracy of 87.01%. In non-transition epochs, per-class F1 scores ranged from 52.71% (N3) to 70.34% (W), while in transition epochs, scores ranged from 62.90% (N1) to 91.77% (REM). Notably, SleepTNet enhances transition epoch detection while maintaining comparable performance on non-transition epochs.