Title: SL-BiLEM: A Structured Learnable Behavior-in-the-Loop Epidemic Modeling Framework for Forecasting and Policy Evaluation

Abstract:

A primary obstacle in epidemic forecasting is the dynamic nature of human behavior, which reacts to disease propagation and generates feedback loops. These loops cause distribution shifts at the moments of policy intervention, thereby compromising the reliability of standard data-driven models when faced with such shifts. To address this, we introduce SL-BiLEM (Structured Learnable Behavior-in-the-Loop Epidemic Model), a framework that utilizes physical constraints as regularization to ensure robust extrapolation capabilities.

The model breaks down effective transmission into the equation $\beta_{\text{eff}}(t,g) = \beta_0(g) \times m_{\text{policy}}(t) \times m_{\text{media}}(t) \times m_{\text{comp}}(t,g)$. By enforcing monotonicity, smoothness, and bounded-jump constraints on the learned compliance function, the model maintains predictive validity even under novel policy regimes. In addition to forecasting, SL-BiLEM supports counterfactual analysis to aid in intervention decision-making.

We validated the model’s forecasting capabilities using three real-world datasets: cruise ship outbreaks, school influenza cases, and school-district COVID-19 surveillance. Furthermore, we assessed counterfactual recovery performance on synthetic benchmarks with established ground truth. The results highlight three key achievements: (1) a 76% performance improvement over neural-mechanistic baselines, coupled with only a 53% degradation in out-of-distribution (OOD) performance compared to an 1142% drop for neural baselines under policy-induced shifts; (2) 100% bootstrap confidence interval coverage across 27 synthetic counterfactual experiments; and (3) Treatment Effect Accuracy surpassing 0.85. These findings position SL-BiLEM as an interpretable resource for public health officials aiming for precise predictions and principled intervention strategies.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC