Graph-aware spatio-temporal attention for forecasting HIV/AIDS case counts in public health surveillance

Provincial-level monthly HIV/AIDS case-count forecasting is important for public health surveillance, early warning, and resource allocation. Existing methods often struggle to capture both cross-regional dependency patterns and complex temporal dynamics from routine notification data.

Using province-by-month case reports from January 2010 to December 2020, a graph-aware spatio-temporal forecasting framework was developed. The model integrates multi-scale 3DCNN encoding, Graph-Aware Spatio-Temporal Attention for modeling inter-provincial relations on a province graph, and a Seasonal Structure-Aware Temporal Module to represent trend and multi-period seasonality.

The proposed method achieved better predictive accuracy and spatial consistency than representative baseline models, with MSE of 12.5336, MAE of 1.4546, SSIM of 0.889, and PSNR of 32.74. Ablation experiments further showed that graph-aware attention and seasonal structure modeling each contributed to performance improvement.

The results indicate that jointly modeling inter-provincial dependency and seasonal temporal structure can improve provincial HIV/AIDS forecasting. This framework provides a useful pathway for more reliable spatial-temporal surveillance and decision support in public health practice.