Predicting comorbidity of metabolic syndrome and daytime sleepiness by machine learning and SHAP value interpretation combined with dietary antioxidants

Objective This thesis aims to develop and validate a machine learning (ML) model that incorporates dietary antioxidants to predict comorbidity of metabolic syndrome (MetS) and daytime sleepiness (DS), and clarify the role of dietary antioxidants in disease prediction. Methods Data were from the 2005–2008 and 2015–2020 National Health and Nutrition Examination Survey database. Antioxidants are selected as key features, including vitamins, minerals, polyphenols, and flavonoids. Additionally, clinical features are incorporated to enhance the model’s performance. The preprocessing of the above features includes removing collinear features, addressing class imbalance, and normaliz​ing data. Within the caret framework in R, 11 ML models are constructed. A systematic method for evaluating and comparing model performance was provided by benchmarking. The Shapley additive explanation (SHAP) value is computed to assess the predictive impact of each feature in the model with the best predictive performance. Results This study included 4840 participants, of whom 842 had Mets and DS comorbidities. ML models were established using 31 antioxidants and 14 clinical features. The AdaBoost model demonstrated a higher prediction accuracy of 83.3%, accompanied by improved precision and recall, as well as higher averages (0.902), and a superior position under the receiver operating characteristic curve (0.825) in the test set. SHAP values analysis revealed that waistline, hypertension, depression, diabetes, epigallocatechin, and vitamin C were the most influential predictors. Conclusion AdaBoost excelled in forecasting the comorbidities associated with Mets and DS. Antioxidants are highlighted as necessary by the SHAP values, with epigallocatechin and vitamin C being the most significant antioxidant factors in this model.