Sleep stage classification using Light Gradient Boost Machine: Exploring feature impact in depressive and healthy participants

Objectives: This study compared Support Vector Machine (SVM), Random Forest, and Light Gradient Boosting Machine (LightGBM) performance in sleep stage classification for patients with depressive disorder and healthy participants. Methods: Participants were divided into Depressive Disorder (DD) and Healthy Control (HC) groups. Overnight polysomnography recordings were obtained, and features were extracted from physiological signals, demographics, and Hamilton Depression Rating Scale (HDRS). The performance of LightGBM, SVM, and Random Forest was conducted, by using HC group (HC/HC case), DD group (DD/DD case), and HC and DD groups (HC/DD case), with 10-fold cross-validation. Results: The classification accuracy of LightGBM for the DD/DD and HC/HC cases (74.78% and 82.12%), respectively, was superior to those of SVM (73.7% and 80.8%) and Random Forest (71.91% and 79.01%). The classification accuracy of LightGBM obtained through HC/DD case (72.25%) was worse than that obtained through DD/DD case (74.78%) to predict sleep stage for subjects with depression. By including demographic features and HDRS score in the analysis, the improvement in the classification accuracy for the DD/DD case (from 74.78% to 76.6%) was greater than that for the HC/HC case (from 82.12% to 82.71%). Feature ranking results revealed that demographic features and HDRS score were the most crucial features for sleep stage classification for the DD/DD case. Conclusions: The classification accuracy of LightGBM for DD group was lower than that for HC group. HDRS score could improve the classification accuracy for DD group. The findings provide insights for managing depressive disorder patients, particularly with sleep-related conditions.