Abstract: Objective  To establish prediction models by using different algorithms for extrapyramidal adverse reactions to antipsychotic drugs in adult hospitalized patients with schizophrenia,and compare them. Methods  Data from the electronic health record system of Beijing Anding Hospital Capital Medical University from 2010 to 2018 was retrospectively collected, and regularization methods to conduct text mining were used to obtain structured clinical data. By using an elastic network model to screen variables that enter the models, 11 algorithms were used to construct prediction models. The sensitivity, specificity, positive predictive value, negative predictive value, area under the curve and decision curve analysis were compared comprehensively. The best performing model was visualized by using Shapley additive explanations to determine the contribution of the prediction factors. Results  The incidence of extrapyramidal adverse reactions to antipsychotic drugs was 41.09%. In the constructed prediction models, CatBoost performed the best (AUC=0.89), and the SHAP summary graph visualized the contributions of predictors. Conclusion  The best model for predicting adverse reactions is CatBoost, which is superior to the latest algorithm TabPFN. The most important risk factors are haloperidol and length of hospital stay.

Key words: antipsychotic drugs, extrapyramidal adverse reactions, prediction model, schizophrenia, hospitalization, electronic health record system, machine learning