291. USING MINOR PHYSICAL ANOMALIES TO PREDICT TREATMENT RESISTANCE IN SCHIZOPHRENIA PATIENTS THROUGH MACHINE LEARNING

Abstract Background Treatment-resistant schizophrenia (TRS) is defined as patients with schizophrenia who despite treatment still have symptoms and poor outcomes. The characteristics of TRS are earlier onset, more severe cognitive deficits, and a poorer outcome. Previous studies showed that neurodevelopmental causes in etiology have a relationship with treatment resistance. In the neurodevelopmental model of schizophrenia, minor physical anomalies (MPAs) were indicated as neurodevelopment markers for schizophrenia. However, the association between MPAs and TRS still needs to be studied. Aims & Objectives The study aimed to utilize machine learning model to construct the risk prediction model to distinguish TRS and non-TRS using MPAs variables. Method The study included 106 TRS patients and 104 non-TRS patients admitted to hospitals in south Taiwan. For selecting the important MPAs variables, we used recursive feature elimination (RFE) to evaluate the important MPAs variables. The logistic regression and two machine learning methods (random forest and support vector machine) were applied to construct the risk prediction models. To interpret the results of risk prediction model, we used SHAP (SHapley Additive exPlanations) values to evaluate each MPAs variable's contribution to the model and utilized summary plot and decision plot to show the results. Moreover, to evaluate the model clinical benefit, we used decision curve analysis to calculate the standardized net benefit and evaluate the model performance for different machine learning models. Results The result showed that using the 10 selected MPAs variables to construct the risk prediction model for discriminating TRS and non-TRS based on logistic regression and two machine learning methods can obtain the Area under the curve (AUC) of 0.67-0.77. The SHAP summary plot and decision plot demonstrated that mouth region and ear region were the important MPAs variables for discriminating TRS and non-TRS. The model performance evaluated by decision curve indicated that random forest model provided the greatest decision-making benefit across clinically relevant threshold probabilities compared to the default intervention strategy of “treat all.” Additionally, we developed prediction models for TRS based on the 10 selected MPAs variables and demographic variables. The model included 10 MPAs variables, sex, and BMI obtained an AUC of 0.67-0.78 based on logistic regression and two machine learning methods. The SVM model demonstrated the greatest predictive ability of the three models (accuracy=0.70, sensitivity=0.71, and specificity=0.68). Discussion & Conclusions In conclusion, MPAs can serve as neurodevelopmental markers of treatment outcomes in schizophrenia. These risk prediction models provide a clinical decision support system for distinguishing TRS from non-TRS and enable early intervention in clinical practice.