Time series forecasting of Valley fever infection in Maricopa County, AZ using LSTM

Abstract

Background

Coccidioidomycosis (CM), also known as Valley fever, is a respiratory infection. Recently, the number of confirmed cases of CM has been increasing. Precisely defining the influential factors and forecasting future infection can assist in public health messaging and treatment decisions.

Methods

We utilized Long Short-Term Memory (LSTM) networks to forecast CM cases, based on the daily pneumonia cases in Maricopa County, Arizona from 2020 to 2022. Besides weather and climate variables, we examined the impact of people's lifestyle change during COVID-19. Factors, including temperature, precipitation, wind speed, PM₁₀ and PM_2.5 concentration, drought, and stringency index, were included in LSTM networks, considering their association with CM prevalence, time-lag effect, and correlation with other factors.

Findings

LSTM can predict CM prevalence with accurate trend and low mean squared error (MSE). We also found a tradeoff between the length of the forecasting period and the performance of the forecasting model. The models with longer forecasting periods have less accurate trends over time and higher MSEs. Two models with different lengths of forecasting periods, 10 days and 30 days, are identified with good prediction.

Interpretation

LSTM algorithms, combined with traditional statistical methods, could help with the forecasting of CM cases. By predicting the CM prevalence, our results can inform researchers, epidemiologists, clinicians, and the public in order to assist public health.

Funding

“Getting to the Source of Arizona's Valley Fever Problem: A Tri-University Collaboration to Map and Characterize the Pathogen Where It Grows” funded by the Arizona Board of Regents.