GeoID-PINN: Identifiability-Aware Regional Epidemic Inference with Geographic Coupling GeoID-PINN：具有地理耦合的可识别性感知区域疫情推断

TL;DR

• GeoID-PINN is a physics-informed neural network combining SIRD epidemic dynamics with spatial regularization to disentangle local transmission from external seeding and reporting effects in regional surveillance data
• A row-stochastic source-composition matrix models spatial dependence, regularized toward spatial priors based on distance, adjacency, commuting, or lead-lag information
• Simulation results show accurate trajectory fitting does not guarantee recovery of the true regional dependence structure, with source-composition error rising from 0.099 (compatible prior) to 0.577 (misspecified prior)
• On 64 Louisiana COVID-19 counties, Forecast-Trained Geo-PINN reduced MSE by 65% (32,957 to 11,468) and MAE by 18% compared to an autoregressive negative-binomial baseline
• County adjacency regularization improved point accuracy but the baseline achieved better distributional fit (lower NLL), highlighting a trade-off between point prediction and probabilistic calibration

Why It Matters

This work addresses a fundamental identifiability challenge in epidemiological modeling: separating local transmission dynamics from external infection pressure and reporting artifacts using only regional surveillance data. For AI practitioners working in public health or spatial modeling, it demonstrates how physics-informed neural networks can encode domain knowledge (SIRD dynamics) while learning spatial coupling structures from data, offering a template for other spatiotemporal inference problems.

Technical Details

• Model architecture: GeoID-PINN integrates a SIRD (Susceptible-Infectious-Recovered-Deceased) compartmental model with a physics-informed neural network, using a row-stochastic source-composition matrix where each row assigns nonnegative weights summing to one across regions
• Spatial regularization: The source-composition matrix is regularized toward spatial priors constructed from distance, adjacency, commuting flows, or lead-lag correlations, enabling structured inference of regional dependence
• Simulation validation: In a four-region simulation with known ground truth, a compatible distance prior achieved source-composition error of 0.099, compared to 0.159 without regularization and 0.577 under strong prior misspecification, while trajectory fit remained similar across conditions
• Real-world evaluation: Applied retrospectively to COVID-19 data from 64 Louisiana counties; Forecast-Trained Geo-PINN reduced MSE from 32,957 to 11,468 and MAE from 70.60 to 57.73 versus an autoregressive negative-binomial baseline, though the baseline achieved lower NLL (5.158 vs 5.346)
• Prior sensitivity: A controlled 15-county comparison showed county adjacency reduced MSE by 6.85% and MAE by 3.1%; similar performance across plausible priors supports structured regularization but not unique edge recovery, necessitating prior-sensitivity and observation-model checks

Industry Insight

• The decoupling of trajectory accuracy from structural identifiability is a critical warning for epidemiological AI: models can produce excellent point forecasts while learning incorrect spatial dependence, meaning deployment in public health contexts requires rigorous prior-sensitivity analysis and uncertainty quantification beyond standard accuracy metrics
• The trade-off between point prediction accuracy (favored by Geo-PINN) and distributional fit (favored by the negative-binomial baseline) suggests hybrid approaches that combine physics-informed structure with probabilistic observation models could yield the best of both worlds for operational epidemic forecasting
• The framework's reliance on spatial priors (distance, adjacency, commuting) that must be specified a priori limits generalizability; future work should explore learned or adaptive priors that reduce human bias in constructing the spatial regularization structure

摘要

GeoID-PINN 是一种物理信息神经网络，将 SIRD 流行病动力学与空间正则化相结合，以从区域监测数据中分离局部传播与外部输入及报告效应
行随机源组成矩阵用于建模空间依赖性，并基于距离、邻接、通勤或领先-滞后信息向空间先验进行正则化
模拟结果表明，准确的轨迹拟合并不能保证恢复真实的区域依赖结构，源组成误差从 0.099（兼容先验）上升至 0.577（误设先验）
在 64 个路易斯安那州 COVID-19 县的数据上，与自回归负二项基线相比，经预测训练的 Geo-PINN 将均方误差（MSE）降低了 65%（从 32,957 降至 11,468），平均绝对误差（MAE）降低了 18%
县邻接正则化提高了点预测精度，但基线模型实现了更好的分布拟合（更低的负对数似然 NLL），凸显了点预测与概率校准之间的权衡

深度分析

一句话总结

• GeoID-PINN 是一种物理信息神经网络，将 SIRD 流行病动力学与空间正则化相结合，以从区域监测数据中分离局部传播与外部输入及报告效应
• 行随机源组成矩阵用于建模空间依赖性，并基于距离、邻接、通勤或领先-滞后信息向空间先验进行正则化
• 模拟结果表明，准确的轨迹拟合并不能保证恢复真实的区域依赖结构，源组成误差从 0.099（兼容先验）上升至 0.577（误设先验）
• 在 64 个路易斯安那州 COVID-19 县的数据上，与自回归负二项基线相比，经预测训练的 Geo-PINN 将均方误差（MSE）降低了 65%（从 32,957 降至 11,468），平均绝对误差（MAE）降低了 18%
• 县邻接正则化提高了点预测精度，但基线模型实现了更好的分布拟合（更低的负对数似然 NLL），凸显了点预测与概率校准之间的权衡

为何重要

本研究解决了流行病学建模中的一个根本性可识别性挑战：仅利用区域监测数据，将局部传播动力学与外部感染压力及报告伪影区分开来。对于从事公共卫生或空间建模的 AI 从业者而言，该工作展示了物理信息神经网络如何在编码领域知识（SIRD 动力学）的同时，从数据中学习空间耦合结构，为其他时空建模任务提供了模板。