Cross-sectional classification of mental health risk indicators in university students using graph convolutional networks: The Atrisk-GCN framework

Identifying students at psychological risk is a growing public-health concern in higher education. Conventional self-report questionnaires are subject to response bias and provide only static snapshots of mental health status. Graph Convolutional Networks (GCNs) offer a principled way to model feature-based relational similarities between individuals, going beyond independent-sample assumptions made by classical classifiers. In this paper we present AtRisk-GCN, a GCN-based framework for cross-sectional classification of questionnaire-derived and text-derived mental-health risk indicators in student populations. Two datasets are studied: (i) a survey-derived dataset of 1 000 records, comprising real medical students from a single institution (Faculty of Biology and Medicine, University of Lausanne, and Lausanne University Hospital, CHUV, Switzerland), drawn from the publicly available data release accompanying Carrard et al. [5]) (Data 1, continuous psychological scale scores as node features) and (ii) a publicly available mental-health text corpus of 27 977 samples (Data 2, TF-IDF representations as node features, no synthetic augmentation). In both cases, edges are constructed from cosine-similarity between node feature vectors using a k-nearest-neighbour (k-NN) scheme; no observed social-network data are available or assumed. On Data 1, AtRisk-GCN achieves an accuracy of 0.861 and an F1-score of 0.857, outperforming LSTM, CNN, and Transformer baselines. On Data 2 the model attains 0.84 accuracy and an F1-score of 0.84, surpassing, under matched input representation, TF-IDFbased Logistic Regression and Linear SVM baselines, and clearly outperforming the graph-free MLP trained on identical feature.