Integrating thyroid function and psychometric profiles for lifetime suicide-attempt risk stratification in bipolar disorder: A multi-algorithm machine-learning study.

Background: Bipolar disorder is a severe mental disorder characterized by recurrent episodes of depression and mania, with a low diagnostic rate. This study aimed to use machine learning methods for risk stratification of lifetime suicide attempts in patients with bipolar disorder based on cross-sectional associations. Methods: The discriminative performance of the models was evaluated using multiple metrics, including the area under the receiver operating characteristic curve (AUC), accuracy, F1-score, precision-recall (PR) curve, Average Precision (AP), and Brier score. LASSO logistic regression was used for variable selection, and SMOTE (synthetic minority over-sampling technique) was applied to handle class imbalance. A sensitivity analysis excluding patients with recent suicide attempts (within 6 months) was performed to reduce reverse causality bias. Thyroid function indicators were detected using chemiluminescent immunoassay with a Roche Cobas e601 analyzer; the reference range for normal TSH was 0.27-4.2 mIU/L, with values outside this range defined as abnormal. Results: We included 1,124 patients diagnosed with bipolar disorder in this study, with a lifetime suicide attempt rate of 31.32%.Among the three models tested, random forests exhibited superior performance metrics, attaining an accuracy (ACC) of 0.938, an AUC of 0.962, and an F1 score of 0.854 compared to gradient boosting (ACC: 0.920; AUC: 0.967; F1 score: 0.828) and the support vector machine (ACC: 0.893; AUC: 0.956; F1 score: 0.808). Suicidal ideation, education level, hopelessness score, the retardation symptoms severity score, and thyroid stimulating hormone levels were identified as the top five predictors. Sensitivity analysis excluding patients with recent suicide attempts (n=98) showed consistent predictor rank-order and maintained RF-AUC = 0.958.Subgroup analysis of euthyroid patients (n=887) preserved the predictor rank and RF-AUC ≥ 0.95. Conclusions: We developed a robust clinical model for predicting (risk stratification of lifetime suicide attempts)suicide attempts in patients with bipolar disorder based on machine learning techniques. This model can assist psychiatric clinicians in understanding suicide risk among individuals diagnosed with bipolar disorder and in identifying those who may require early intervention or preventive measures. We identified a set of 20 clinical and biological markers significantly associated with lifetime suicide attempts in bipolar disorder patients. Importantly, due to the cross-sectional design, these associations cannot establish temporal precedence or support prospective risk prediction. A key limitation is that suicide attempts were assessed retrospectively without a defined time window between predictor measurement and outcome occurrence. Future prospective validation in multi-site, multinational cohorts is required to confirm the model's clinical utility.