Simulation Study on Digital Divide Propagation Path of Elderly Medical Treatment Based on Graph Neural Network and Public Social Network Data

Abstract

To address the self-reinforcing diffusion of the digital divide in elderly medical treatment and the difficulty of precise intervention, this study proposes a Propagation–Adoption Coupled Graph Neural Network (PAC-GNN) and a path-level interpretable simulation framework. Using three types of public social network data, a heterogeneous multi-layer network with four node types and four edge layers was constructed, retaining 191607 users and 6057580 interaction records, with elderly users accounting for 19.3%; the elderly identification classifier achieved an AUC of 0.91. PAC-GNN integrates dual-channel decoupling, random-walk path encoding, cross-layer attention, and a causal masking mechanism to enable fine-grained modeling of propagation paths. Experimental results show that compared with the suboptimal baseline EvolveGCN, PAC-GNN reduces state prediction RMSE by 14.2%, improves path prediction F1 by 18.3%, and lowers literacy-dimension error by 19.5%; under 20% node perturbation, its performance decay is 8.3 percentage points lower than that of baselines. Simulation identifies five typical propagation patterns, among which intra-community diffusion accounts for 28.4% with an elderly participation rate of 68.7%, while cross-community bridging paths contribute nearly 60% of global state change. Multi-objective optimization indicates that the community-bridging strategy achieves the best overall balance among effectiveness, cost, and benefit equity.