Spatio-temporal prediction and macro-level correlates of hand injuries in manufacturing regions: a multi-source data fusion study using graph neural networks and its implications for targeted prevention

Abstract

Abstract Background Hand injuries impose a heavy burden on manufacturing-intensive regions, yet the spatio-temporal patterns of regional risk and their macro-level correlates remain poorly understood. Most existing studies focus on individual-level risk factors and rely on single data sources, leading to “one-size-fits-all” prevention strategies that are often inefficient. This study aims to fill this gap by integrating multi-source data and a spatio-temporal graph neural network (STGNN) to identify high-risk areas and key associated factors. Methods We developed a prediction framework based on an STGNN that fuses injury surveillance records, meteorological data, nighttime light intensity, and socioeconomic indicators from Hunan Province, China (2011–2023, N = 182). A fused adjacency matrix combining geographical proximity and economic linkages was constructed to capture spatial dependencies. Model performance was evaluated using RMSE, MAE, and R² against SARIMA, geographically weighted regression, Random Forest, XGBoost, DLinear, GCN, and a spatio-temporal CAR model benchmarks. Interpretability was assessed via SHapley Additive exPlanations (SHAP) and attention weight visualization. Results The STGNN model achieved superior predictive performance compared to all benchmark models. SHAP analysis identified proportion of secondary industry and nighttime light intensity as the two most important positive correlates. High-risk areas persistently clustered in the manufacturing-intensive Changsha-Zhuzhou-Xiangtan city cluster. Conclusions Our study demonstrates that industrialization level and economic activity intensity are the core macro-level correlates of hand injury risk as recorded by hospital data. The STGNN framework provides a tool for accurate spatio-temporal risk prediction, enabling spatially targeted resource allocation and regional joint prevention in manufacturing-intensive regions. The integration of this predictive intelligence into public health planning may facilitate a shift from reactive treatment to proactive, precision prevention. All findings are based on predictive associations and do not imply causation.