Construction and validation of a predictive model for suicidal attempts in patients with mood disorders after discharge based on machine learning methods: A two-year follow-up study.

BACKGROUND: Suicide attempts (SA) in patients with mood disorders (MD) should be paid enough attention. Currently, there is a lack of relevant research on the prediction of SA of inpatients with MD in psychiatric wards after discharge. We tried to construct a SA prediction model for hospitalized patients within two years of discharge through a variety of machine learning algorithms, and verify the relevant performance of the model. METHODS: We followed up 1099 patients for two years, a total of 865 patients completed. In order to avoid the influence of redundant features, we used the LASSO regression method to reduce the dimension, and used eight machine learning algorithms to construct the SA prediction model. At the same time, we evaluated the machine learning model with the best stability based on sensitivity, AUC and FI scores. Then we used the local interpretation technique of SHapley Additive ExPlanations (SHAP) to analyze the contribution of each feature to SA within two years after discharge. RESULTS: We tried to analyze the prediction model based on the characteristics of LASSO regression. The SA prediction model constructed by random forest (RF) method showed good robustness (testing AUC = 0.770, training AUC = 0.944). SHAP interpretation technique found that SA occurred within 3 months was the most important feature, and previous NSSI was more indicative than previous SA and suicidal behavior (SB). CONCLUSION: The RF method can well construct a risk prediction model for SA in hospitalized MD patients within two years of discharge.