Predicting depressive symptoms among Chinese college students using recurrent neural networks with longitudinal data

AI Summary

Recurrent neural network models accurately predicted depressive symptoms (AUC 0.921-0.952), outperforming traditional techniques and improving sensitivity over longer follow-up.
Family relationship problems predicted early-stage symptoms; childhood trauma and preexisting mental health had lasting impact; sleep and self-rated health consistently signalled risk.
Early identification enables targeted interventions: family-based support, trauma-informed counselling and sleep health education to prevent depression and promote student wellbeing.

Sci Rep. 2026 Jun 1. doi: 10.1038/s41598-026-54695-7. Online ahead of print.

ABSTRACT

College students face a higher risk of depression than their non-college peers. However, the predictors of depressive symptoms among college students and their relative importance remain inconclusive. This study aimed to develop predictive models for depressive symptoms among Chinese undergraduate students using a four-year longitudinal dataset and to explore the importance ranking of predictive factors. A cohort of 1,500 college students from Sichuan University was recruited and followed annually for four years. Computer-based questionnaires were used to collect demographic, behavioral and psychological data. Using a recurrent neural network (RNN), four predictive models were constructed to capture the temporal dynamics of depression risk as students progressed through college. The prevalence of depressive symptoms ranged from 3.60% to 6.27% across the four waves. All RNN models demonstrated strong predictive performance (AUC: 0.921-0.952), with sensitivity improving over longer follow-up periods, outperforming traditional models, such as multilevel linear regression, k-nearest neighbor and support vector machines. Analysis of feature importance revealed that family relationship issues were key predictors of early-stage depressive symptoms. Childhood traumatic experiences and preexisting mental health conditions had lasting impacts, while health-related factors, such as subjective sleep quality, sleep disorders, and self-rated health, were consistent indicators across years. These findings have important practical implications for university mental health programs, suggesting that early identification of high-risk students could enable targeted interventions such as family-based support, trauma-informed counseling, and sleep health education tailored to students’ developmental stages. Such data-driven strategies may improve prevention efforts and promote student well-being.

PMID:42219415 | DOI:10.1038/s41598-026-54695-7

Document this CPD