Application of Physics-Informed Neural Networks to Asthma Epidemiology

Abstract

Asthma is a chronic lung condition in which the airways become inflamed and narrow, and overproduce mucus, making breathing difficult. Asthma affects 262 million people worldwide, and as exposure to pollution is a key risk factor in developing conditions, increasing urbanization is often accompanied by an increase in asthma prevalence, particularly is lesser-developed regions. To examine the relationship between asthma and pollution, we constructed a system of ordinary differential equations using a compartmental model with susceptible, exposed, and infected components, as well as a pollutants component to act as a pathogen. The parameters of the model that describe the relationship between components are difficult to measure and are currently unavailable. Thus, a physics-informed neural network approach was taken to compute parameter values for a given dataset. The network was trained on artificially generated time series data of each component. The error on estimated parameters was calculated and optimized based on known input parameters to the system. Once the error is reduced via hyperparameter tuning, the network can compute accurate parameters given real-world time series data to allow for realistic modeling of asthma-pollution epidemiology.