Enhancing automatic sleep stage classification with cerebellar EEG and machine learning techniques

Sleep disorders have become a significant health concern in modern society. To investigate and diagnose sleep disorders, sleep analysis has emerged as the primary research method. Conventional polysomnography primarily relies on cerebral electroencephalography (EEG) and electromyography (EMG) for sleep stage scoring, but manual scoring is time-consuming and subjective. This study investigated the potential application of cerebellar EEG combined with machine learning in automatic sleep stage classification. Twenty-five male mice underwent 24-h cerebral EEG/cerebellar EEG/EMG recording, and manual sleep staging was performed. Various machine learning models, including Light Gradient Boosting (LGBoost), Extreme Gradient Boosting, Categorical Boosting, Support Vector Machine, Logistic Regression, Random Forest, Long Short-Term Memory and Convolutional Neural Network, were applied for automatic sleep stage classification. The performance of different models and the efficacy of cerebellar EEG, cerebral EEG, and EMG were compared under different training:test set ratios. Cerebellar EEG exhibited significant differences in power spectral density across wakefulness, non-rapid eye movement sleep stages, and rapid eye movement sleep stages, particularly at frequencies >7 Hz. LGBoost, Extreme Gradient Boosting, and Categorical Boosting models showed comparable performance, with LGBoost being selected for further analyses due to its shorter computation time. Cerebral EEG consistently demonstrated the highest precision, recall/sensitivity, and specificity in classifying sleep stages across all training:test set ratios, followed by cerebellar EEG, which outperformed EMG. Combining the top 5 cerebellar EEG features with cerebral EEG features yielded better classification performance than combining EMG features with cerebral EEG features. Using the top 20 features, the model achieved mean area under the receiver operating characteristic curve values of 0.98 ± 0.08, 0.98 ± 0.10, and 0.99 ± 0.07 for wakefulness, non-rapid eye movement sleep stages, and rapid eye movement sleep stages, respectively. The cerebellum may play a unique and important role in sleep-wake regulation. Incorporating cerebellar EEG into polysomnography has the potential to enhance the accuracy and efficiency of sleep stage classification.