University students in Lebanon are facing war superimposed on a prolonged national crisis, creating an urgent need for scalable tools to detect mental health problems early. We used machine learning (ML) to predict depression, anxiety, and perceived stress among students during the 2024 war in Lebanon, compare algorithms, identify key predictors, and assess how clinical cut-offs influence performance. We conducted a cross-sectional online survey among 225 students at the American University of Beirut. Outcomes were measured using PHQ-9, GAD-7, and PSS-10 with moderate and severe thresholds. Predictors included sociodemographic, lifestyle, conflict-related, and psychosocial variables. Nine supervised algorithms were trained using stratified train–test splits, cross-validated tuning, and balanced versus SMOTE-enhanced pipelines. Models were selected primarily using test-set PR-AUC, with recall as a secondary criterion, and were additionally evaluated using ROC-AUC, Brier scores, and calibration curves. Symptoms were highly prevalent: 61.3% screened positive for moderate-to-severe depression, 59.6% for anxiety, and 73.3% for stress; 27–35% met severe thresholds. Balanced Random Forest performed best for depression and anxiety (PR-AUC 0.93 and 0.88), and balanced calibrated AdaBoost for stress (PR-AUC 0.95), with improved calibration for anxiety and stress. Key predictors included declines in healthy eating, sleep, and physical activity; greater social media use; higher avoidance and lower approach coping; larger household size; and fear for personal and family safety. These exploratory findings highlight the promise of interpretable ML for advancing early mental health risk detection in conflict-affected universities, provided that future studies confirm performance in larger and independent samples. (Access to the Article)