Explaining complex dynamical systems using conditional SHAP analysis with application to multi-variant epidemic dynamics.

Understanding and explaining the dynamics of complex systems is a critical yet challenging task. Such a challenge often arises from high dimensionality, nonlinearity, and randomness, which together make it difficult to study the key drivers influencing system behavior. One illustrative example is infectious disease dynamics. Outbreaks like COVID-19 have major public-health impacts, and timely vaccines and non-pharmaceutical interventions can lessen them. However, COVID-19, influenza, and other pathogens continually generate new strains, so standard low-dimensional compartmental models can be uninformative or misleading (a complexity comparable to many biological models). Policymakers therefore need tools to understand and identify key drivers of the system in such high-dimensional settings with random events. In this study, we address interpretability of such complex models through a machine learning framework. Specifically, we introduce a novel, multi-strain, multi-vaccine compartmental model and analyze it using AI methods and conditional SHapley Additive exPlanations importance measures. By training a surrogate neural network, we efficiently approximate the dynamics and perform feature importance analysis. Unlike traditional approaches, our conditional importance analysis reveals how a feature’s influence varies with other features, capturing key interactions among features in disease dynamics. While our case study focuses on epidemiological systems, the proposed framework offers a general methodology for understanding the drivers of complex nonlinear dynamical systems.