Predicting Sleep Quality Using Hybrid Stacking Ensemble Of Random Forest And Catboost With Shap Interpretability

Abstract

Sleep quality is an important indicator of physical and mental well-being, yet its prediction from lifestyle variables remains challenging because behavioral and physiological factors interact nonlinearly. This study develops a hybrid stacking ensemble that combines Random Forest and CatBoost as base learners and Logistic Regression as the meta-classifier for three-class sleep-quality prediction (Poor, Moderate, Good). The Sleep Health and Lifestyle Dataset contains 374 records, of which only 109 are unique, and is preprocessed through blood-pressure decomposition, derived-feature construction, categorical encoding, and numerical standardization. Models are evaluated with Accuracy, Precision, Recall, F1-Score, and multiclass ROC-AUC under Stratified 10-Fold Cross Validation, both in its standard form and in a duplicate-aware form in which identical records are kept in the same fold. Under the duplicate-aware protocol, Random Forest achieved 97.86% Accuracy, 91.06% macro F1-Score, and 0.9970 ROC-AUC; CatBoost achieved 97.33%, 90.70%, and 0.9930; and Hybrid Stacking achieved 97.06%, 90.51%, and 0.9784. Recall for the minority Poor class (n = 12) was 66.7% for all three models, and stacking did not improve on its base learners. SHAP analysis identifies Stress Level and Sleep Duration as the two dominant features. The results are specific to this small, largely duplicated dataset and do not support clinical use.