A Random Forest Model for Predicting Social Functional Improvement in Chinese Patients with Schizophrenia After 3 Months of Atypical Antipsychotic Monopharmacy: A Cohort Study

PURPOSE: Impaired social functions contribute to the burden of schizophrenia patients and their families, but predictive tools of social functioning prognosis and specific factors are undefined in Chinese clinical practice. This article explores a machine learning tool to identify whether patients will achieve significant social functional improvement after 3 months of atypical antipsychotic monopharmacy and finds the defined risk factors using a multicenter clinical study. PATIENTS AND METHODS: A multicenter study on atypical antipsychotic (AAP) treatment in Chinese patients with schizophrenia (SALT-C) was conducted from July 2011 to August 2018. Data from 550 patients with AAP monopharmacy from their baseline to their 3-month follow-up were used to establish machine learning tools after screening. The positive outcome was an increase in the Personal and Social Performance (PSP) scale score by ≥10 points. The predictors were a range of investigator-rated assessments on symptoms, functioning, the safety of AAPs and illness history. The Least Absolute Shrinkage and Selection Operator (LASSO) was used for the feature screening and ranking of the predicted variables. The random forest algorithm and five-fold cross-validation for optimizing the model were selected to ensure the generalizability and precision. RESULTS: There were 137 patients (mean [SD] age, 41.1 [16.8] years; 77 [58.8%] female) who had a good social functional prognosis. A lower PSP score, taking a mood stabilizer, a high total Positive and Negative Symptom Scale (PANSS) and PANSS general subscale score, unemployment, a hepatic injury with medication, comorbid cardiovascular disease and being male predicted poor PSP outcomes. The generalizability of the PSP predictive tool was estimated with the precision–recall curve (accuracy of 79.5%, negative predictive value of 92.6% and positive predictive value of 57.1%) and receiver operating characteristic curve (ROC) (specificity of 81.8% and sensitivity of 78.7%). CONCLUSION: The machine learning tool established using our current real-world data could assist in predicting PSP outcome by several clinical factors.