Climate variability is increasingly associated with sanitation breakdowns and subsequent outbreaks of climate sensitive diseases in tropical regions, creating a need for forecasting tools that operate at local scales. Existing indices and conventional learning pipelines often underuse spatial connectivity and long-range temporal dependence, which limits their ability to anticipate cascades following extreme rainfall, waterlogging, and sewer overflow. We propose ClimaSan-GNN, a spatiotemporal graph model that couples GRU-based temporal encoding with attention-guided message passing to predict standardized sanitation and public health risk scores from ERA5-Land climate drivers and derived vulnerability indicators. The dataset spans a graph of N geospatial nodes observed daily for 1,827 days (2020–2024), supporting region-specific and cross-region evaluation. Unlike T-GCN, GWN, and DCRNN, ClimaSan-GNN uses attention to adapt neighbor influence under event- driven conditions and incorporates a robustness-oriented training design with uncertainty-aware outputs to flag low-confidence forecasts during extremes and partial data. Across major tropical regions, the model achieves RMSE = 0.112, MAE = 0.083, and R2 = 0.938, consistently improving over strong spatiotemporal baselines. These forecasts can drive municipal early warning dashboards and guide flood-period sanitation actions such as targeted desludging, chlorination, and resource allocation. The findings support evidence-based planning aligned with SDG 3, SDG 6, SDG 11, and SDG 13.