Optimization of Health Service Utilization Among Elderly People with Chronic Diseases in Rural Ethnic Minorities in Northwest Yunnan Using Graph Neural Networks

Abstract

The demand for health services among the elderly with chronic diseases in rural ethnic minority areas in northwest Yunnan continues to rise, but there is a significant imbalance in the use of services. Existing studies have mostly focused on disease combinations, ignoring the temporal and geographical differences in medical behavior, making it difficult to accurately identify high-risk and service-mismatched populations. This paper applies GNN (Graph Neural Network) to construct a heterogeneous graph of patients, institutions, and service behaviors, model the dynamic path of individuals in the service system, identify high-frequency and lost contacts, and help optimize resources and service equity. This paper constructs a heterogeneous health service utilization graph network composed of patients, medical institutions, and geographical units and integrates medical records and spatial proximity to form a graph structure; HAN (Heterogeneous Graph Attention Network) is used to embed node features and extract behavioral similarities and service path characteristics; GAT (Graph Attention Network) classifier is trained based on the embedding results to identify frequent medical treatment and potential loss groups; further applies geographic and social variables such as ethnicity, terrain, and road accessibility to enhance the model's sensitivity to regional differences; finally, service optimization suggestions are generated by combining the centrality and path distribution in the graph to assist in the precise allocation of health resources and path reconstruction. The model dynamically analyzes the edge load and node betweenness centrality in the patient-institution service path, drives resource optimization decisions, and specifically guides the spatial reconstruction of the addition of mobile medical points and cross-regional collaborative nodes. The experiment shows that there are significant spatial and structural differences in medical services in northwest Yunnan: the accessibility index of Diqing Prefecture is 68 minutes (only 29 minutes in Dali), and traffic barriers are prominent in ethnic minority plateau areas; the matching rate of multiple disease combinations (3+) is as low as 68.6%, but the utilization rate is as high as 1.138, reflecting the imbalance of complex comorbidity resources and matching; the coverage index of mountain unit G18 is as low as 0.31, and the Macro-F1 of the model in this paper reaches 0.83, which is better than the traditional model XGBoost (0.76), which can effectively identify high-risk groups, locate service bottlenecks and optimize resource allocation.