Scalable prediction of suicidal risk in university students: a three steps machine learning approach in university settings

Abstract

Suicide is a leading cause of death among young adults, with university students representing a particularly vulnerable subgroup. Although prevention efforts often depend on psychopathological symptom assessments, these are resource-intensive and less practical for large-scale screening. This study evaluated whether easily collectable contextual and psychosocial variables can independently predict suicidal risk in university students. Data from 470 university students were analyzed using random forest machine learning (ML) algorithm. Three models were developed: Model 1 included 53 contextual and psychosocial variables (sociodemographic, lifestyle, social support, academic stressors, and health); Model 2 relied solely on 10 psychopathological symptom scales; and Model 3 integrated all variables. Predictive performance improved progressively across the three models (AUC = 0.72, 0.77, and 0.80, respectively). Key predictors included both contextual and psychosocial variables (e.g. low social support, diminished university belonging) and psychometric evaluation of symptoms (anxiety, depression, coping strategies). Taken together, these results show that contextual and psychosocial variables provide meaningful information for identifying university students at risk for suicide, and psychometric symptom measures offer further refinement of risk detection. Findings are discussed in relation to the Interpersonal Theory of Suicide and Escape Theory. This ML approach offers a scalable framework for suicide risk screening in academic settings.