Speaker
Description
The Susceptible–Infectious–Recovered (SIR) model has long been a foundational framework for modeling infectious diseases and has played an important role in informing public health policies during the COVID-19 pandemic. However, traditional SIR-based approaches primarily rely on epidemiological case data and often fail to account for behavioral and societal factors that influence disease transmission dynamics.
In this study, we propose a novel framework that integrates large language models (LLMs) with physics-informed neural networks (PINNs) to enhance epidemic modeling. Specifically, LLMs are employed to derive structured social signals from large-scale textual sources, capturing information related to policy interventions, public risk perception, and healthcare system pressure. These signals are incorporated into a physics-informed neural network framework to estimate the time-varying transmission rate governing the SIR dynamics.
By embedding the governing differential equations of the SIR model directly into the neural network training process, the proposed approach ensures consistency with epidemiological dynamics while leveraging both observed case data and LLM-derived social signals. Applying the framework to COVID-19 case data from Seoul, we demonstrate that integrating LLM-based social signals improves the model’s ability to capture temporal variations in transmission dynamics compared with conventional SIR models and standard PINN-based approaches.
