Construction and Verification of a Risk Prediction Model for Suicidal Ideation in Patients With Bipolar Disorder: A Machine Learning Analysis

Background: Bipolar disorder (BD) is closely associated with suicidal ideation (SI). The development of an effective prediction model for SI in BD patients could facilitate early risk identification in high-risk groups. Methods: This study employed a cross-sectional design. Patients with BD were enrolled from three tertiary hospitals between July 2021 and July 2024. All participants were randomly allocated to training (n = 204) or testing (n = 88) sets at a 7:3 ratio. A hybrid feature selection strategy integrating the data-driven Boruta algorithm with clinical expertise was used to identify potential predictors of SI. Nine machine learning algorithms were applied to the training set to construct SI prediction models. The optimal model was selected through comprehensive evaluation of the area under the receiver operating characteristic curve (AUC), F1 score, balanced accuracy, sensitivity, and other indicators. SHapley Additive exPlanations (SHAP) analysis was used to rank and interpret the importance of features in the best-performing model and to assess their contributions to SI. Results: A total of 292 patients with BD were analyzed, of whom 149 (51.03%) reported SI during the past week. Among the nine models, the random forest (RF) model demonstrated superior predictive performance, with an AUC of 0.915 (95% CI: 0.850–0.965), a balanced accuracy of 0.818, a sensitivity of 0.891, a specificity of 0.833, a precision of 0.826, a average precision of 0.922, an F1 score of 0.860, and a Matthews correlation coefficient of 0.704. The SHAP analysis revealed that quality of life was the most influential predictor, followed by the number of depressive episodes, history of suicide attempts, cognitive functioning, and emotional abuse in childhood trauma. Conclusions: RF-based models can effectively predict SI in BD patients and inform clinically targeted interventions.