Decoding olfactory-depression-appetite pathways in nursing-home elderly: a knowledge multi-relational graph attention...

The growing application of AI-based clinical systems underscores the need for analytical frameworks to interpret sensory-affective patterns in elderly populations. To solve this problem, the Knowledge Multi-Relational Graph Attention Generative Neural Network (KMRG-GANN) is proposed as an AI-assisted sensory-psychological model that incorporates symbolic knowledge. The framework uses the Olfactory_Appetite_NursingHomes_Dataset, which comprises 193 records of participants with olfactory detection identification measurements, log-scaled gustatory thresholds, cognitive analyses, depressive indicators, anthropometry, and appetite-nutrition outcomes measured using standardized sensory and clinical protocols. The Dual Selective Fusion Transformer Network (DSFTN) is a model that builds cross-domain feature embeddings and enables selective fusion across the sensory-cognitive and physiological-nutritional domains. The KMRG-GANN combines a Knowledge Graph and Graph Neural Network (KG-GNN) with Multi-Relation Graph Attention Networks (MRGAT), which are trained with the Fisher Mantis Optimizer (FMO), to produce symbolic, relation-sensitive pathways that reflect the progression from olfactory deterioration to increased depressive symptoms and loss of appetite. The experimental analysis shows that the predictive accuracy is 99.9%, higher than transformer-based, graph-based, and attention-only baselines. The mediation analysis shows that depression is the carrier of 56.8% of the total olfactory effect on appetite control in the elderly of nursing homes.