Biomarker signatures as predictors of future impulsivity in schizophrenia: a multi-center study

Siqi Liu¹Yixiao Chen²Lei Zhang¹Xu Zhang³Jiali Min¹Yaqin Yang⁴Manru Li¹Zheya Cai³Yanwei Sun¹Jiayi Wang³Ruyi Shui³Zhihao Chen⁵Hui Li⁶Fazhan Chen⁷Jiaojiao Hou⁸Guoquan Zhou¹Enzhao Zhu^3*

• ¹Shanghai Putuo District Mental Health Center, Shanghai, China
• ²Shanghai Tongji Hospital, Shanghai, China
• ³Medical School, Tongji University, Shanghai, China
• ⁴Shanghai Yangpu District Mental Health Center, Shanghai, China
• ⁵East China University of Science and Technology, Shanghai, China
• ⁶Shanghai Mental Health Center, Shanghai, China
• ⁷Shanghai Pudong New Area Mental Health Center, Shanghai, China
• ⁸Aachen University, Aachen, Germany

While clinical scales for impulsivity assessment in psychiatric settings are widely used, evidence linking laboratory biomarkers to impulsivity remains limited. This study evaluated the prognostic value of routinely collected biomarkers for future impulsivity risk and developed a machine learning-based prediction model. We analyzed data from 1,496 first-admission schizophrenia (SCZ) patients across four specialized psychiatric hospitals (2016)(2017)(2018)(2019)(2020)(2021)(2022)(2023). A total of 99 features, including 91 routinely tested biomarker measurements, four treatment-related indicators, and four demographic or psychometric variables, were evaluated. Impulsivity was assessed using the Impulsive Behavior Risk Assessment Scale within one week of admission. Five machine learning models were trained with 10-fold cross-validation (n=993) and externally validated in an independent cohort (n=503). Model performance was assessed using the area under the receiver operating characteristic curve (AUROC), and biomarker importance was evaluated using SHapley Additive exPlanations (SHAP). Of 1,496 SCZ patients, 882 (59.0%) exhibited high impulsivity. CatBoost outperformed other models, achieving an AUROC of 0.749 in cross-validation and 0.719 in external testing. SHAP values identified key biomarkers, revealing heterogeneous response patterns for uric acid (UA), globulin (GLO), apolipoprotein E (APOE), and others. Combining biomarkers with clinical data improved prediction, increasing AUROC from 0.652 to 0.749 in cross-validation and from 0.655 to 0.721 in external testing. Subgroup analyses revealed sex-specific patterns, with exploratory analysis suggesting sex-modified relationships between UA and impulsivity. These findings highlight the utility of routine biomarkers for early identification of high-risk individuals with SCZ and suggest the importance of incorporating sex-specific factors in predictive modeling.