DUA-HIR: A Multi-Modal Dual-Attention Deep Learning Framework for Personalized Health Intervention

To address the personalized recommendation needs of health interventions for college students, existing methods still have limitations in accurately matching individual health conditions with intervention effects. This study proposes a Dual-attention Utility and Interest Representation (DUA-HIR) model. Based on a deep learning framework, the model jointly learns user health features, historical behavior sequences, and intervention utility through an embedding layer, multi-layer fully connected networks, and Long Short-Term Memory (LSTM) sequence modeling, enabling dynamic modeling of intervention utility and user interests. It introduces a multimodal fusion mechanism that combines wearable device physiological data, psychological and stress indicators with historical behavior sequences to generate dynamic interest vectors, thereby capturing the time-varying characteristics of interests and their changes with physiological and psychological states. The DUA-HIR constructs a multi-task learning framework consisting of a representation layer and a prediction layer: the representation layer extracts user health needs, intervention utility, and interest features. The prediction layer evaluates the potential health improvement effects and adoption probability of intervention schemes for target users. Experiments are validated based on FatSecret platform data and small-sample data of college students. Ablation experiments show that the dual-attention mechanism and multimodal dynamic interest representation significantly enhance feature weighting and interest matching capabilities. Compared with baseline models, the DUA-HIR achieves AUC (Area Under Curve), Recall@1, and Recall@5 of 0.931, 0.889, and 0.933 respectively, representing improvements of 0.22%, 0.34%, and 0.43% over the second-best model. DUA-HIR also performs optimally in terms of average weight loss success rate, weight loss magnitude, and user satisfaction. The results demonstrate that the model effectively improves the scientificity and personalization of health intervention recommendations for college students, providing data-driven intelligent decision support for artificial intelligence-based health interventions.