Abstract
# Processed dataset and source code for STGNN‑HTR: Hand trauma risk prediction across 14 cities in Hunan, China (2011–2023) ## OverviewThis repository contains the complete source code, preprocessing scripts, and processed dataset for STGNN‑HTR, a reproducible spatio-temporal graph neural network framework for multi-source health risk forecasting. ## Study Area14 prefecture-level administrative divisions in Hunan Province, China: Changsha, Zhuzhou, Xiangtan, Yueyang, Changde, Hengyang, Chenzhou, Yiyang, Loudi, Yongzhou, Shaoyang, Huaihua, Xiangxi, Zhangjiajie ## Temporal Coverage2011–2023 (13 years, annual data) ## Variables**Target variable:** Hand trauma incidence rate (per 100,000 population) **Feature variables (9 total):** Resident population, population density, road network density, manufacturing/construction employee ratio, secondary industry GDP share, GDP per capita, physician density (per 1,000 population), college education proportion, and hand trauma incidence rate from previous year (autoregressive) ## Data Preprocessing- Missing value imputation: Linear interpolation for college education proportion- Normalization: Min-max scaling to [0,1] ## Data Split (Chronological)- Training: 2011–2019 (9 years)- Validation: 2020–2021 (2 years)- Testing: 2022–2023 (2 years) ## Model Architecture- Hybrid graph construction (geographic adjacency + economic similarity)- 2-layer Graph Convolutional Network (GCN) for spatial dependencies- Gated Recurrent Unit (GRU) for temporal dynamics- Multi-head attention (4 heads) for adaptive feature fusion ## UsagePlease see the GitHub repository for complete installation and usage instructions: https://github.com/CIAM-Lab/stgnn-hand-injury ## CitationIf you use this code or dataset, please cite:Chen Y, Qin X, Chen S. (2026). STGNN‑HTR: A reproducible spatio-temporal graph neural network framework for multi-source hand trauma risk prediction. PeerJ Computer Science. ## LicenseMIT License