Speaker
Description
Individuals adapt to epidemics, creating a feedback loop between disease spread and human behavior. Incorporating this dynamic into infectious disease models is critical for evaluating interventions against future pandemic threats. We embedded an XGBoost algorithm, trained on real-world survey data (Imperial College London YouGov COVID-19 Behavior Tracker; Netherlands, Jun 2020–Jan 2021; N=5,611), directly into an agent-based model to predict contact restricting behavior. Agent attributes (age, gender, perceived severity, willingness to isolate, working outside the home) combined with population-level hospitalizations drove behavioral predictions, with adopted behavior reducing agent contacts (edge-masking; 47% reduction). The model was parameterized to wildtype COVID-19 (R0=3.28) and run on three synthetic networks (small world, random, preferential attachment; 10,000 nodes; 100 simulations). Additional mechanisms captured local epidemic awareness spread, behavioral fatigue, and prevalence-dependent re-adoption. Behavior was least effective at slowing epidemic spread on preferential attachment and random networks (median durations: 87 and 145 days; final sizes: 65–70%) versus baseline (87–92%), while the small world network showed substantial curve flattening (403 days; final size: 71%). Integrating machine learning algorithms trained on real data into agent-based models offers a principled framework for modeling behavioral feedback and evaluating pandemic interventions.
