Abstract: Objective To explore the characteristics of two machine learning algorithms in the construction of suicidal behavior predictive models for bipolar disorder inpatients, and to compare their performance so as to provide a basis for the prevention and control of suicidal behavior of inpatients with bipolar disorder. Methods Clinical data of bipolar disorder inpatients treated during Jan. 2010 and Dec. 2017 were retrospectively analyzed. After the influencing factors of suicidal behavior were screened with Chi-square test analysis, Adaboost and binary Logistic regression were used to construct two suicidal behavior classification models. The characteristics of the two models were compared with recall ratio, precision ratio and F1 value. Results A total of 7,782 bipolar disorder inpatients were enrolled, among whom 1,661 had suicidal ideation or attempted suicide and the rate of suicidal behavior was 21%. Between the two models, Adaboost performed better. Diagnosis subtype and past suicide history were the two most important risk factors in both models. Conclusion There are only slight differences between the two models based on different machine learning algorithms, both having a low average performance. More factors are needed to improve the model performance. Current depression episode or mixed state and past suicide history are the most predictive traits for suicidal behavior in bipolar disorder inpatients. Preventive measures should be taken to address suicidal behavior risks accordingly.

Key words: Machine learning, Suicidal behavior, Influening model, Bipolar disorder, Adaboost, Logistic regression