COS 220-3 - Predicting outbreak dynamics from early-stage data

Wednesday, August 9, 2023

4:00 PM – 4:15 PM PDT

Location: B113

Presenting Author(s)

NL

Nathan C. Layman, PhD

Research Scientist
EcoHealth Alliance, Washington, United States

Coauthor(s)

NR

Noam Ross, PhD

Principal Scientist
EcoHealth Alliance, New York, United States
EM

Emma Mendelsohn

Senior Data Scientist
EcoHealth Alliance, New York, United States

Abstract: Epidemics are challenging to forecast in their early stages because of limited observed case counts and lack of information on disease properties, which depend not only on the pathogen itself but also on environmental, demographic, and socioeconomic contexts. Yet decision-makers require guidance in just these data-scarce periods. The goal of the study was to employ hierarchical approaches to epidemic forecasting that draw on information from previous outbreaks with related pathogens or with similar outbreak source, pathogen traits, geography, host population or environmental conditions. To address this, we assembled and curated a database of 2,238 epidemic time series across a variety of human and veterinary diseases, and developed a series of hierarchical models that predict both nonparametric epidemic curves and mechanistic parameters of compartmental models based on similarity to previous outbreaks and early-stage case counts. We apply this method to predict near-term case counts and to forecast the overall duration, size and shape of outbreaks. In addition, we also identify the degree that each pathogen, host and environmental factor captures information capable of informing future outbreaks. For example, we found that by transferring knowledge about how outbreaks are shaped by access to health care, rural population size, and climatic indicators, our method was able to increase the accuracy of malaria forecasts – even when trained on only four malaria time-series. This approach holds the potential to increase epidemic preparedness by providing early-stage epidemic risk assessment and to guide surveillance efforts in maximizing disease forecast accuracy. We show that these methods are effective and can greatly improve early-stage forecasting when compared to approaches that consider each disease in isolation or fail to include ecologically relevant predictors.