The Role of Cognitive Function in Predicting Metabolic Risk in Schizophrenia: A Multi-Model Comparison Incorporating Clinical Features

Rui LI^1,2Xuan Ren¹Tingyun Jiang^2*Jiawen Huo^1,2Junjiao Ping²Shuyi Zhu¹Aoxiang LUO^1*

• ¹Guangdong Pharmaceutical University, Guangzhou, China
• ²Third People's Hospital of Zhongshan, Zhongshan, China

Background: Patients with schizophrenia frequently exhibit metabolic abnormalities that are closely associated with cognitive impairment. However, clinically applicable risk-stratification tools based on concise and generalizable indicators remain limited. This study evaluated the predictive value of cognitive and clinical features for metabolic risk stratification and compared the discriminative performance of traditional statistical and machine-learning models. Methods: In this cross-sectional study, 213 patients with schizophrenia who received treatment at Zhongshan Third People's Hospital between September 2024 and September 2025 were enrolled according to the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5). Based on the diagnostic criteria for metabolic syndrome in the Chinese Guideline for the Prevention and Treatment of Type 2 Diabetes (2017 edition), patients were categorized into three groups: High-risk, Critical, and MS. General clinical data, symptom ratings, and cognitive assessments (Chinese version of the MATRICS Consensus Cognitive Battery, MCCB) were collected. Features were selected using the Boruta algorithm and screened for multicollinearity, followed by the construction of multinomial logistic regression, random forest (RF), extreme gradient boosting (XGBoost), and support vector machine (SVM) models; class imbalance was addressed using SMOTE. Results: Years of education, processing speed, verbal learning, visual learning, and reasoning/problem solving demonstrated stable and independent predictive contributions across models. Age, age at onset, and negative symptoms were also retained during feature selection. The RF model achieved the best overall discriminative performance (macro-average AUC = 0.789; Macro-F1 = 0.603), whereas the SVM model showed superior performance in identifying minority classes (balanced accuracy = 0.725; Macro-F1 = 0.625). These results remained consistent after controlling for clinical symptoms and general demographic variables. Conclusions: Modeling based on concise clinical and cognitive indicators can effectively achieve metabolic risk stratification in patients with schizophrenia. Rather than relying on a single algorithm, combining the complementary strengths of RF and SVM and selecting models according to specific clinical needs and data characteristics may improve the identification of high-risk individuals and support proactive intervention and management.