Health Risk Assessment in Smart Elderly Care Using Multimodal Data Fusion

This study addresses the growing demand for advanced computational frameworks for personalized health risk estimation, a critical problem in applied machine learning and intelligent healthcare systems. Existing approaches, while effective in controlled settings, face persistent challenges in handling multimodal data heterogeneity, missing observations, and limited interpretability, which restrict their applicability in real-world elderly care environments. To overcome these limitations, we propose a two-stage framework comprising a graph-based temporal encoder, VitalGraph-Net, and a domain-aware reasoning module, RARE. VitalGraph-Net learns structured temporal representations from heterogeneous sensor streams by modeling inter-modal correlations and long-term dependencies through dynamic graph construction and attention-enhanced recurrent modeling. RARE further enhances interpretability by integrating semantic priors, uncertainty estimation, and contextual reasoning to support reliable real-time decision-making. Experiments conducted on four public multimodal elderly care datasets demonstrate consistent performance gains over state-of-the-art methods, including up to a 12.4% improvement in prediction accuracy and notable increases in AUC and F1-score across diverse deployment scenarios, while reducing computational overhead by 34%. These results indicate that the proposed framework achieves robust, accurate, and interpretable health risk prediction across heterogeneous sensing environments. By jointly leveraging temporal analytics and domain-aware reasoning, this work advances the development of resilient and actionable machine learning models for intelligent elderly care systems.