Machine learning for sleep stage classification in control and pilocarpine-induced epileptic rats

Abstract

Purpose To develop a robust methodology for automatic classification of sleep stages in rats using machine learning (ML) models, validated for control rats and those with pilocarpine-induced epilepsy, and to compare the relative importance of the features for the models trained separately for each group of rats. Methods EEG and EMG recordings from 12 male Wistar rats (6 with epilepsy) were used. Visual classification of sleep stages was performed by researchers, and features were extracted from the signals. Features were normalized using Quantile Transformation (QT) and used to train ML models, including Multilayer Perceptron (MLP), Support Vector Machine (SVM), Random Forest (RF) and Balanced Random Forest (BRF). Four recordings were used for training and two for testing, with hyperparameters tuned via Bayesian Optimization. Feature importance was assessed using Permutation Feature Importance (PFI). Results The MLP model was selected for the Control group, with a macro F1-score of 89.91%. For the Epilepsy group, MLP was also selected, with a macro F1-score of 81.88%. The models demonstrated good classification metrics on test data, with F1-scores of at least 82.64% in the Control group and 70.99% in the Epilepsy group. PFI revealed significant differences in feature importance between the Control and Epilepsy groups, with the model of the Control group relying more on EEG sleep waves (gamma, theta and alpha), and the model of the Epilepsy group relying more on EEG sleep wave variations and EMG features. Conclusions ML models effectively classified sleep stages in rats, showing good generalization to new data and human evaluators. The QT and PFI techniques enhanced model interpretability, suggesting potential applications in other rat groups and future validation in humans.