Enhancing the objective identification of depression risk in university students via a hybrid intelligent multimodal assessment model

The objective assessment of mental health among university students is a critical prerequisite for effective early intervention, yet traditional diagnostic methodologies are frequently constrained by the inherent subjectivity of self-reporting. To overcome these limitations, this study proposes a novel hybrid intelligent assessment framework designed to extract and synthesize complex emotional features from multimodal data. The core architecture integrates Convolutional Neural Networks (CNNs) with Self-Attention Mechanisms (SAMs), enabling the simultaneous capture of localized spatial features and global contextual dependencies. To optimize the balance between computational efficiency and predictive accuracy, we introduce a residual-based Bottleneck Transformer architecture, systematically replacing high-layer convolutional operators with self-attention layers to facilitate long-range feature modeling. The model was evaluated for depression risk score prediction using behavioural video-derived optical-flow features and structured psychological assessment data. Empirical results show that the proposed hybrid framework achieves lower prediction error and higher explained variance than the evaluated baseline models under the same experimental protocol. These findings suggest that the integration of transformer-based architectures provides a robust, objective foundation for large-scale, automated mental health screening support. By enabling timely, data-driven risk flagging for professional follow-up, this research offers a scalable technical solution for improving psychological outcomes within higher education environments.