Predictive value of biomarkers for short‑term impulsivity risk in hospitalized patients with schizophrenia: a machine learning study

Impulsivity is a common and critical issue in hospitalized schizophrenia patients. This study aimed to examine the predictive value of objective biomarkers and to accurately predict impulsivity based on machine learning. A retrospective cohort study was conducted using data from four mental health institutions in Shanghai. Five machine learning models were trained to predict future impulsivity based on routinely tested biomarkers and clinical variables. Model performance was primarily evaluated using the area under the receiver operating characteristic curve (AUROC). Clinical utility was assessed by net benefit at Youden’s best cut-off. Feature importance was quantified using SHAP analyses. A total of 2595 first-admission patients with schizophrenia were included. On the external test cohort, the XGBoost model achieved the best predictive performance, with an AUROC of 0.847. Biomarkers such as β₂-microglobulin, d -dimer, creatinine, creatine kinase, and glycated hemoglobin emerged as important predictors of impulsivity. Incorporating these biomarkers into the models improved both predictive performance and clinical utility. This study shows that machine learning models can predict the short-term risk of impulsivity in hospitalized patients with schizophrenia using routinely collected clinical and biomarker data. Adding objective biomarkers improved predictive performance, especially when symptom data were unavailable.