Spatio-Temporal Graph Networks for Infectious Disease Forecasting

This direction is split into other directions: exogenous infection pressure, spatial independence assumption. Its last version stays readable and citable here.

Version 1 of 1246 papers296 conceptsfrozen 2026-10-04

Traditional epidemiological models often treat geographic regions—cities, counties, or countries—as isolated silos. In a standard compartmental model or a simple time-series forecast, the number of cases in London is predicted primarily based on the history of cases in London. However, pathogens do not respect administrative boundaries. They travel along the arteries of human mobility: flight paths, commuting routes, and social interactions.

Recent research has shifted toward a network-centric view of disease spread. By representing geographic units as nodes in a graph and human movement as edges, researchers can capture the "spillover" effects that drive localized outbreaks.

The Failure of the Independence Assumption

The fundamental limitation of traditional forecasting is the assumption of spatial independence. As noted by Benjamin Jeffrey, David M. Aanensen, and Nicholas J. Croucher in their work on antibiotic resistance, ignoring neighboring regions leads to sensitive, inaccurate predictions. When an outbreak occurs in a "hotspot," it acts as a leading indicator for connected "cold" regions.

This connectivity is rarely just about physical proximity. While Wei Liu and colleagues found that population density and hospital proximity drove the initial spread in Wuhan, long-distance transmission is better explained by non-Euclidean networks. Dongxia Wu and Rose Yu argue in DeepGLEAM that standard grid-based models fail because they cannot account for the irregular, long-range dependencies of global travel.

Modeling Mobility as Message Passing

To solve this, researchers are increasingly using Graph Neural Networks (GNNs). In this framework, each node (a region) has a state representing its current disease burden. The edges (mobility) allow for "message passing," where the infection state of one node influences its neighbors.

George Panagopoulos, Giannis Nikolentzos, and Michalis Vazirgiannis demonstrated this in their Transfer Graph Neural Network framework. By combining GNNs with meta-learning (MAML), they could "transfer" the dynamics learned from a mature outbreak in one country to a country in the early stages of a pandemic. The model learns the underlying logic of how mobility translates to transmission, allowing it to make accurate predictions even when local data is sparse.

Similarly, the County Augmented Transformer (CAT) developed by Shihao Yang and Tuo Zhao uses a multi-scale approach. It augments state-level forecasts with granular county-level data, recognizing that the "macro" trend is an aggregate of "micro" interactions across a spatial network.

Dynamic Networks and Latent Influence

A key insight in this field is that the network itself is not static. Stephanie Armbruster showed that the optimal network structure for forecasting shifts based on human behavior. During lockdowns, sparse economic networks best explain the data; when movement is unrestricted, dense connectivity is required.

Sometimes, the transmission pathways are not explicitly recorded in travel surveys. Minkyoung Kim developed the Latent Influence Point Process (LIPP) to uncover these "invisible" causal links. By fusing disease data with digital footprints from social media, the model can quantify the "feedback loops" between regions that drive sustained outbreaks.

Hybridizing Physics and Data

While deep learning models like GNNs are powerful at pattern recognition, they can lack the "common sense" of biological reality. This has led to the rise of hybrid models. DeepGLEAM combines mechanistic simulations (which follow the "first principles" of biology) with deep learning (which corrects for the biases and simplifications of those principles).

This hybridization allows models to respect the known laws of epidemiology—such as the way a virus evolves as it moves between sub-populations, as explored by Wenxian Shi and Regina Barzilay in their work on viral evolution—while still leveraging the massive, unstructured datasets of the modern world, from Google Search Trends to satellite imagery.

By treating the world as an interconnected graph rather than a collection of independent points, these models provide a more realistic and predictive map of how threats—whether they be viruses, antibiotic resistance, or even ambulance demand—propagate through space and time.

Papers in this version

Showing 60 of 246 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_spatio_temporal_graph_networks_for_infec_956963d2,
  title = {Spatio-Temporal Graph Networks for Infectious Disease Forecasting},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/spatio-temporal-graph-networks-for-infectious-disease-forecasting/rdr_956963d2e91f4dfca17505f6f02ad9e5},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_spatio_temporal_graph_networks_for_infec_956963d2_v1,
  title = {Spatio-Temporal Graph Networks for Infectious Disease Forecasting},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/spatio-temporal-graph-networks-for-infectious-disease-forecasting/rdr_956963d2e91f4dfca17505f6f02ad9e5/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}