JMIR Publications

ABSTRACT

Background:

Mental health challenges are prevalent among Canadian higher-education students, with significant rates of depression and anxiety often going untreated, due to reduced early detection, stigmatizing beliefs, and practical barriers. The U-Flourish longitudinal electronic survey study launched in 2018 engages new cohorts of incoming undergraduate students and repeatedly collects data about mental health and well-being and access to support.

Objective:

U-Flourish survey data provides a unique opportunity to train evidence-based prediction risk models and a personalized recommendation engine to signpost students to indicated mental health support based on their own data.

Methods:

Two approaches were used and integrated in developing the risk prediction and recommendation engine: (i) clinically-defined rules by experts in the field to detect current and predict the risk of future anxiety and depression and to signpost students to appropriate care using a stepped care approach and based on clinical factors (i.e., self-harm and suicidal thoughts, symptom levels, lifetime history); (ii) Machine Learning (ML) models, trained with additional data including family history, early adversity, and stress indicators, to predict future risks of clinically significant depression (PHQ-9) and anxiety (GAD-7). Models were created using the XGBoost algorithm, and a 70:30 ratio for training and testing with 10-fold cross-validation.

Results:

27.5% of students at entry to university from 2018-2023 were identified as having potentially clinically significant levels of anxiety and depression and signposted to university mental health services based on the clinically-defined rules. Optimizing thresholds to reduce false negatives, the ML models predicted anxiety and depression over the year in students screening negative at baseline with accuracy comparable to reported clinical screening as evidenced by sensitivity equal to or above 90% for all models trained). Models had high negative predictive value (89% or above), balanced against low specificity. Individuals identified at risk for anxiety or depression were signposted primarily to self-guided resources supporting proactive prevention. Model findings also demonstrated that abbreviated screens (PHQ-2, GAD-2), with potential to reduce respondent burden and improve adherence, can be used without compromising sensitivity. Indeed, PHQ-2 displayed a 90% sensitivity and GAD-2 a 92% sensitivity. SHAP analyses revealed other predictive factors including childhood trauma, family history of mental illness, and functional impairment associated with reported depression and anxiety symptoms.

Conclusions:

The risk prediction models and recommendation engine’s dual approach rationalizes support allocation and promotes targeted early intervention and prevention, potentially improving capacity to address the increasing burden on university mental health services. Future directions include further refinement based on a larger harmonized and enriched dataset, independent validation and implementation studies to estimate the complex factors that influence uptake, reach to services and acceptability across more diverse student users.