Student depression has a measurable effect on academic performance, social life and wellbeing, yet conventional diagnosis is slow and subjective. This study tested whether it can be predicted from routine data instead.
The Problem
In Nepali higher education, student mental health remains an under-researched area. Many students experience academic pressure, sleep disruption, financial stress and social isolation — yet few seek formal support. Early detection could help institutions intervene before a crisis.
Methodology
Using a dataset of approximately 27,900 student records covering demographic, academic and lifestyle variables, we compared three classification models:
- Logistic Regression — baseline binary classifier (accuracy: 0.75)
- Random Forest — ensemble method (accuracy: 0.73)
- Gradient Boosting — sequential ensemble (accuracy: 0.75)
Key Findings
All three models performed in a similar band around the mid-0.70s, but the feature analysis proved the more actionable result:
- Academic pressure emerged as the single strongest predictor of depression
- Sleep duration was the second most influential feature
- Financial stress and family relationships contributed meaningfully but less strongly
The finding that academic pressure and sleep are the strongest predictors suggests that institutional interventions — workload management, sleep hygiene campaigns, flexible assessment — could have a measurable impact on student wellbeing.
Implications
These results suggest that routine institutional data can serve as an early-warning system. Rather than relying solely on self-reporting or clinical assessment, colleges could use predictive models to identify at-risk students and provide targeted support.
Limitations & Future Work
The dataset was cross-sectional, limiting causal inference. Future work will explore longitudinal models and the integration of real-time data from learning management systems. The fairness implications of predictive models in educational contexts also warrant further investigation.