<p>Machine learning (ML) offers promise for suicide risk stratification in depressed youth, yet its clinical application remains methodologically challenging. Using prospective data from 602 Chinese patients aged 15–24 years collected between January 2022 and June 2023, we developed ML models to predict suicide attempts within 30 days after treatment. From 102 clinical and psychosocial predictors, only 30 suicide attempts (5.0%) were observed, resulting in a limited predictor-to-event ratio. Seven algorithms were trained on 70% of the sample (<i>n</i> = 421; 21 events) using 10‑fold cross‑validation and tested on the remaining 30% (<i>n</i> = 181; 9 events), with model selection emphasizing regularization and parsimony to reduce overfitting risk. Among the algorithms, the Support Vector Machine (AUC = 0.831) and Elastic Net (AUC = 0.811) achieved the best test performance, while more complex models such as random forests and deep learning exhibited poor generalization. A combined SVM + EN ensemble reached an AUC of 0.84 in cross‑validation and identified a high‑risk decile with a 20% suicide attempt rate compared to 3.6% among remaining patients (RR = 5.53), although confidence intervals were wide due to the small number of events. These findings demonstrate the technical feasibility of ML‑based short‑term risk stratification but also underscore important methodological constraints. When retrained using only 15 LASSO-selected predictors, the model’s discrimination remained comparable (AUC = 0.82), supporting robustness against over-fitting. Low event counts limited model stability, cohort homogeneity and single‑country recruitment restricted generalizability, and the lack of temporal validation precluded assessment of model drift. Consequently, the models presented here should be viewed as proof‑of‑concept rather than evidence of clinical readiness, providing an empirical basis for future validation in larger and more diverse longitudinal cohorts.</p>