IHBEM: Data-driven integration of behavior change interventions into epidemiological models using equation learning

Given the antigenic characteristics of a virus, human behavior is the single most important determinant of disease transmission. Human behaviors relevant to disease spread such as social distancing, wearing face coverings, or testing when asymptomatic depend on a host of factors including risk perceptions, physical ability as well as the availability of resources and opportunities. Policy interventions by health agencies or other decision makers can impact these factors to alter human behaviors. Using decision models to tailor these interventions by time and sub-population can ensure efficiency (e.g., low cost), effectiveness (e.g., less hospitalizations), and equity (e.g., fairness in access to pharmaceuticals). The overall goal of this project is to incorporate behavior change driven by public health interventions into mathematical epidemiological models to inform decision making and policy evaluation during infectious disease outbreaks. The investigators consider respiratory diseases in general, and use COVID-19 as an example to validate the approach and quantify impact. The proposed methods can be generalized to other applications where policy makers target behavior change, such as medication adherence. In Aim 1, the investigators will trace the impact of policy interventions on infection-preventive behaviors through mechanisms of action (i.e., capability, opportunity, and motivation). Nine types of policy interventions will be considered (education, persuasion, incentives, coercion, restriction, training, nudging, modeling, and enablement) in relation to two types of preventive behavior – interpersonal protection (i.e., social distancing, wearing a face mask) and service utilization (i.e., testing, vaccination). The empirical work involves a dynamic meta-analysis of interventions to reduce the spread of COVID-19, supplemented by Delphi methods. The investigators will develop an online tool that will enable researchers to contribute to the meta-analysis and use the resultant weighted-average effect sizes as inputs for agent-based modeling. The results of Aim 1 will be operationalized by integrating adaptive human behaviors into an agent-based model (ABM). However, realistic ABMs with a large number of agent types and complex behavioral and social processes are computationally intensive to simulate, analytically intractable, and may not be generalizable. These drawbacks may inhibit the comprehensive analysis and validation of ABMs and thereby prevent their utilization for decision- and policy-making during a pandemic. Thus, in Aim 2, the investigators propose an equation learning framework to derive ordinary differential equation (ODE) models from ABMs. The investigators also introduce novel regularization techniques that incorporate biophysical constraints to provide interpretable results for decision-makers. These ODE models and the learned functional forms approximating the impact of interventions on behavioral and social processes that drive disease spread will be used in Aim 3 to inform policies through bilevel optimization models. This project is jointly funded by the Division of Mathematical Sciences (DMS) in the Directorate of Mathematical and Physical Sciences (MPS) and the Division of Social and Economic Sciences (SES) in the Directorate of Social, Behavioral and Economic Sciences (SBE). This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.