Can machine learning spot student distress in wartime?

The signal

During Lebanon’s 2024 war, researchers surveyed 225 students at the American University of Beirut and tested nine supervised machine-learning approaches against screening measures for depression, anxiety and perceived stress.

The strongest internal models performed well on the study’s held-out test data: balanced Random Forest led for depression and anxiety, while calibrated AdaBoost led for stress. But the paper is careful about what that means. This was a cross-sectional study at one university, and the authors describe the findings as exploratory rather than a validated clinical tool.

The psychological reading

The interesting part is not simply that a model can classify distress. It is what becomes legible as signal.

Among the important predictors were changes in sleep, healthy eating and physical activity; social-media use; coping style; household size; and fear for personal or family safety. In a conflict setting, context is not statistical noise around an individual mind. It is part of the psychological environment the model is learning from.

That matters because prediction can easily acquire the aura of diagnosis. Here, the outcomes came from established screening instruments — PHQ-9, GAD-7 and PSS-10 — but screening status is not the same thing as a clinician establishing a diagnosis.

The evidence boundary

The paper reports high internal predictive performance in this sample, but it does not establish that the same models would perform equally well at another university, in another conflict, or in routine care.

The study used one institution, a relatively small sample and cross-sectional data. Model performance therefore needs external validation, calibration checks and scrutiny for false positives and false negatives before any operational use. The authors themselves call for larger, independent samples.

What PSYTECH would watch next

The next question is not whether the model can produce a risk score. It is whether a validated system could improve access to support without converting uncertainty into false clinical certainty.

Useful follow-up evidence would include external validation, subgroup performance, calibration over time, clinician and student acceptability, and whether prediction changes outcomes rather than merely identifying risk more efficiently.

Sources

- Hoteit, R., Bou-Hamad, I. & El Morr, C. Machine learning prediction of depression, anxiety, and stress among university students during wartime in Lebanon. Scientific Reports (2026).