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

Researchers introduce SL-BiLEM, a machine learning framework that improves epidemic forecasting by accounting for how human behavior changes in response to disease spread and policy interventions. The model uses physical constraints to maintain accuracy even when facing novel policy scenarios, demonstrating 76% improvement over existing neural baselines and potential applications for public health decision-making.

SL-BiLEM addresses a critical limitation in epidemic modeling: the assumption that historical patterns remain constant. Traditional data-driven models fail catastrophically when policies change because human behavior shifts unpredictably—people alter mobility, contact patterns, and compliance based on perceived risk. This paper proposes decomposing transmission into interpretable components (baseline transmission, policy effects, media influence, and compliance) with mathematical constraints that enforce realistic behavior.

The broader context reflects growing recognition that pure neural networks, while powerful pattern-matchers, lack the structural knowledge necessary for robust predictions beyond training distributions. Public health agencies have repeatedly experienced model failures when lockdowns or vaccination campaigns fundamentally altered population dynamics. SL-BiLEM bridges machine learning and mechanistic modeling by embedding domain knowledge as regularization.

For public health infrastructure and policy makers, this framework offers tangible value. The model achieves credible uncertainty quantification (100% bootstrap coverage) and demonstrates treatment effect accuracy exceeding 0.85, enabling evidence-based intervention planning rather than guesswork. The 53% out-of-distribution degradation versus 1142% for neural baselines represents a meaningful difference in decision-quality for resource allocation during crises.

Looking forward, this work signals broader trends in AI: hybrid architectures combining learned flexibility with physical constraints are outperforming pure deep learning for scientific applications. As regulatory bodies demand interpretability and robustness from AI systems in critical domains, methods like SL-BiLEM become foundational infrastructure rather than niche research.

• →SL-BiLEM achieves 76% forecasting improvement over neural baselines with dramatically better robustness to policy-induced distribution shifts.
• →The framework decomposes transmission into interpretable, constrained components that maintain validity under novel policy regimes.
• →Model demonstrates 100% bootstrap confidence interval coverage on synthetic counterfactuals and treatment effect accuracy above 0.85.
• →Hybrid physics-informed machine learning approaches outperform pure neural networks for epidemic forecasting in high-stakes public health contexts.
• →Framework enables principled intervention planning and counterfactual analysis for public health decision-makers.