New AI model GeoID-PINN improves epidemic tracking with 65% error reduction

Researchers Weixiong Hua and Fan Bu from Washington University introduced **GeoID-PINN** (Geographic Identifiability-Aware Physics-Informed Neural Network), a novel AI model designed to improve regional epidemic inference by explicitly modeling geographic dependencies in disease spread.

GeoID-PINN combines physics-informed neural networks (PINNs) with a row-stochastic source-composition matrix to represent spatial transmission dynamics. Unlike traditional models that treat regions in isolation, this approach incorporates geographic priors—such as distance, adjacency, commuting, or lead-lag patterns—into the neural network’s regularization. In a controlled simulation with four regions, the model achieved a source-composition error of just 0.099 when using a compatible geographic prior, compared to 0.159 without regularization and 0.577 with a misspecified prior—highlighting the importance of structured spatial modeling.

When evaluated on real COVID-19 data from 64 Louisiana counties, GeoID-PINN (in a forecast-trained variant) reduced mean squared error (MSE) from 32,957 to 11,468 and mean absolute error (MAE) from 70.60 to 57.73 versus an autoregressive negative-binomial baseline. While the baseline had better negative log-likelihood (indicating superior distributional fit), GeoID-PINN delivered significantly more accurate point predictions. Additional experiments showed that incorporating county adjacency reduced MSE by 6.85% and MAE by 3.1% in a 15-county subset, suggesting that even simple geographic priors can substantially improve forecasting performance.