Spatially Directional Dual-Attention GAT for Spatial Fluoride Health Risk Modeling

TL;DR

SDD-GAT presents a spatial graph neural network that decouples geographic proximity from semantic similarity via a dual-graph design and augments message passing with directional attention to capture anisotropic spatial processes. A spatial smoothness regularization term enforces local coherence, yielding geographically smoother and more reliable risk predictions for fluoride exposure and fluorosis across Guizhou. Empirical results show superior regression and classification accuracy, higher Moran’s I indicating spatial coherence, and robust generalization under distribution shifts and data perturbations, all with practical CPU-only deployment feasibility. The work offers a generalizable framework for fine-grained, geospatial health risk modeling that can be extended to spatio-temporal and multimodal settings for scalable public health decision support.

Abstract

Environmental exposure to fluoride is a major public health concern, particularly in regions with naturally elevated fluoride concentrations. Accurate modeling of fluoride-related health risks, such as dental fluorosis, requires spatially aware learning frameworks capable of capturing both geographic and semantic heterogeneity. In this work, we propose Spatially Directional Dual-Attention Graph Attention Network (SDD-GAT), a novel spatial graph neural network designed for fine-grained health risk prediction. SDD-GAT introduces a dual-graph architecture that disentangles geographic proximity and attribute similarity, and incorporates a directional attention mechanism that explicitly encodes spatial orientation and distance into the message passing process. To further enhance spatial coherence, we introduce a spatial smoothness regularization term that enforces consistency in predictions across neighboring locations. We evaluate SDD-GAT on a large-scale dataset covering over 50,000 fluoride monitoring samples and fluorosis records across Guizhou Province, China. Results show that SDD-GAT significantly outperforms traditional models and state-of-the-art GNNs in both regression and classification tasks, while also exhibiting improved spatial autocorrelation as measured by Moran's I. Our framework provides a generalizable foundation for spatial health risk modeling and geospatial learning under complex environmental settings.

Figures (8)

• Figure 1: The model ingests geospatially distributed fluoride-related features (e.g., fluoride concentration, pH value, soil type, and location coordinates) and constructs two complementary graphs: a spatial graph based on geographic proximity and a feature graph based on semantic similarity. Each graph is processed by a directional attention module that encodes pairwise angular and distance information. The outputs of both branches are fused and passed to dual output heads for regression (Dean’s Fluorosis Index) and classification (fluoride risk level). This design effectively captures anisotropic spatial dependencies and latent feature relationships, enhancing both predictive performance and spatial coherence.
• Figure 2: Regression results on test set. Scatter plot comparing predicted and ground-truth DFI values. The SDD-GAT model aligns closely with the 1:1 diagonal, demonstrating strong regression accuracy.
• Figure 3: Confusion matrix for binary fluorosis risk classification. The SDD-GAT model exhibits high discriminative performance with strong recall and precision.
• Figure 4: Robustness evaluation under noisy and incomplete input conditions. Top row: Performance degradation under increasing Gaussian noise. Bottom row: Impact of feature dropout. SDD-GAT demonstrates superior robustness across all metrics compared to baseline models.
• Figure 5: Computational efficiency comparison across models. Left: average training time per epoch. Middle: inference latency per batch of 128 samples. Right: total trainable parameters.