Explainable spatiotemporal graph learning for early identification of localized infectious disease risk
Year: 2026
Type: article
Abstract: Localized infectious disease forecasting is important for targeted public health interventions, but remains challenging because of complex spatiotemporal dynamics and limited model interpretability. We propose an explainable spatiotemporal graph learning framework for region-level infectious disease forecasting, using COVID-19 in Sweden as a case study. The framework integrates demographic, socioeconomic, syndromic surveillance, and epidemiological data within a nationwide graph of fine-grained geographic areas. We introduce an adaptive gating mechanism that modulates graph-derived spatial representations using temporal context while the geographic adjacency graph remains fixed. To improve interpretability, we combine analysis of gating behavior with post hoc explainable AI analyses of temporal, structural, and spatial feature sensitivities. The framework is evaluated across 3-, 7-, 14-, and 21-day forecasting horizons under both offline (held-out temporal split) and online (prequential predict-before-update) learning settings and compared with LSTM, tree-based machine-learning models, and standard and advanced spatiotemporal graph models. In the offline setting, the spatiotemporal graph convolutional network with adaptive gating (STGCN-AG) generally performed better than LSTM, tree-based, and standard spatiotemporal graph baselines, with the clearest gains at shorter horizons. Among the advanced graph forecasting models evaluated, STGCN-AG performed best at the 3-day horizon and remained broadly competitive at longer horizons. Nevertheless, in the offline evaluation, all models showed limited absolute predictive performance, which further deteriorated as the forecasting horizon increased, highlighting the difficulty of longer-term localized forecasting. In the online setting, the models continued to track evolving incidence under sequential updating, although predictions at longer horizons showed greater temporal lag and weaker responsiveness to rapid changes in observed incidence. The adaptive gate indicates when, where, and across which latent dimensions the model more strongly modulates graph-derived representations. These findings highlight both the potential and the remaining challenges of explainable spatiotemporal graph learning for localized infectious disease forecasting.
Source: npj Digital Public Health
Institutions: Lund University, Karlstad University, Skåne University Hospital
Language: English
Subfield: Epidemiology
Field: Medicine
Domain: Health Sciences
Open Access status: hybrid
Awards: G8030113496, G9001396927, G9001397494
OpenAlex ID: W7236290775
