Multimodal spatiotemporal graph convolutional attention network for dynamic risk stratification and intervention strategy generation in rare disease rehabilitation nursing

Rare disease rehabilitation nursing presents unique challenges due to heterogeneous clinical manifestations, limited sample sizes, and complex comorbidity patterns that render traditional risk assessment tools inadequate. This study proposes a novel multimodal spatiotemporal graph convolutional attention network (MSTGCA-Net) for dynamic risk stratification and intervention strategy generation in rare disease rehabilitation. The framework integrates four principal innovations: a heterogeneous patient relationship graph construction scheme encoding clinical similarities, an adaptive multimodal fusion module employing cross-attention mechanisms, a spatiotemporal encoder capturing both inter-patient relationships and longitudinal dependencies, and a knowledge-guided intervention generation component. Experiments conducted on a retrospective cohort of 2,847 patients with 156 rare disease categories demonstrate that MSTGCA-Net achieves superior performance compared to baseline methods, with accuracy of 0.867, F1 score of 0.845, and AUC of 0.923. Expert evaluation of generated intervention strategies yielded favorable assessments across clinical appropriateness, safety, and feasibility dimensions. The attention-based architecture provides interpretable predictions that facilitate clinical adoption. This framework offers promising decision support tools for precision rehabilitation nursing in rare disease populations.