Forecasting CWD with Physics-Informed GNNs

Chronic wasting disease is a fatal prion disease in deer, elk, and moose with no known treatment. Surveillance data is sparse, zero-inflated, and spatially heterogeneous (most county-years report zero cases because no one tested) which breaks both naive regression and pure mechanistic SIR modeling.

Our team built a county graph (1,438 nodes, 8,064 edges) from a 31k-record surveillance dataset across 16 U.S. states, then trained a spatio-temporal GNN with a physics-informed loss that combines data fit with a one-step Metapopulation-SIR residual.

Architecture

Spatio-temporal GNN: static and time-varying county features pass through stacked GATv2 blocks and a GRU, producing three heads (infection, harvest, SIR β/γ) trained jointly under a weighted-MSE + one-step SIR residual + regularization hybrid loss. Untested counties are masked out of the infection loss.

The encoder fuses static features (lat/lon/area) and time-varying features (positive/negative tests, harvest, facilities) through GATv2 attention blocks and a GRU. Three heads decode infections, harvests, and the per-county SIR parameters $\beta, \gamma$ . The physics term enforces that consecutive infection predictions roughly satisfy the SIR update $I_t \approx I_{t-1} + (\beta S_t I_t - \gamma I_t),$ regularizing the GNN toward epidemiologically plausible trajectories without forcing a fixed ODE form.

Forecast — 2020–2021 season

Ground truth	Predicted (GNN + harvest + SIR loss)

The full hybrid model hit $\text{MAE} = 0.73$ on held-out 2020–2021 infections, outperforming $\text{AR}(1)$ (1.08) and an ODE-only Metapopulation-SIR baseline (2.62), while keeping interpretable SIR parameters at the county level.