Cross sectional discrimination of lifetime suicide attempt history in a Chinese survey targeting men who have sex with men and transgender people

Psychosocial measures and infection markers may help characterize self-reported suicide attempt history in samples recruited from sexual and gender minority communities. Cross-sectional associations cannot establish whether those measures precede an attempt. We examined internally validated discrimination of lifetime attempt history rather than future suicide risk. We analysed a sample of 2,275 participants from three Chinese study sites; 560 (24.6%) reported lifetime suicide attempt history on a single item. Forty-five candidate variables covered demographics, psychosocial measures, sexual behaviour and four infection markers plus an any-positive composite. Logistic regression, random forest and XGBoost were evaluated using identical stratified five-fold splits, with preprocessing fitted within training folds. We assessed paired model differences, calibration, sensitivity analyses and geographic internal-external validation. Mean AUCs were 0.817, 0.836 and 0.833 for logistic regression, random forest and XGBoost, respectively. At a probability threshold of 0.5, sensitivities were 0.414, 0.352 and 0.471. Ten repetitions of paired five-fold cross-validation did not establish a model difference after Holm correction (all adjusted P values > = 0.073). ACE count had the largest mean absolute held-out SHAP value. The infection variables had relatively low SHAP contributions. Models excluding PHQ-9 and GAD-7 had AUCs of 0.814–0.835. The syphilis by continuous PHQ-9 interaction was not supported (likelihood-ratio P = 0.711; parametric-bootstrap P = 0.719). This three-city study characterizes psychosocial differences associated with reported lifetime suicide attempt history, with childhood adversity contributing prominently to model discrimination. The persistence of discrimination without recent depression and anxiety scores supports broader psychosocial investigation. These findings inform community-based mental health research and the design of prospective studies; the models require further evaluation before use for screening or prediction of subsequent attempts.