Rainfall-Adaptive Graph Transformer Framework with Federated Edge Deployment for Multi-Label Waterborne Disease Risk Classification

Abstract Waterborne diseases have become a continuous problem across India’s groundwater-dependent states. Predicting them has become a challenge because water’s chemical composition data sits scattered across concerned departments of state, and privacy regulations prevent sharing. This paper introduces the Biochemical Weather Groundwater Network (BWG-Net), a federated deep learning framework that brings intelligence to where the data exist and removes the need to transfer sensitive records to a central server. The framework is trained on 11,240 Central Ground Water Board station records from five Indian states, combining 16 biochemical water quality parameters with 6 meteorological variables fetched via station coordinates from the India Meteorological Department archives. It classifies contamination into 7 health risk categories based on World Health Organization and Bureau of Indian Standards limits. The architecture combines a rainfall-adaptive dynamic graph, which updates how stations are connected based on precipitation intensity, with a cross-modal attention module that learns how weather conditions amplify the danger of specific chemical contaminants. Federated learning lets each state train on its own data and share only model weights. On held-out test data, BWG-Net achieved 90.4% accuracy and an area under the receiver operating characteristic curve (AUC-ROC) of 0.939, outperforming extreme gradient boosting (XGBoost) and a convolutional neural network–long short-term memory (CNN-LSTM) model by 16.6 and 9.5 AUC-ROC points, respectively. The system also runs on a Raspberry Pi 4 at 73 ms per inference, making it deployable in field settings.