Forecasting new diseases in low-data settings using transfer learning

Epidemic models fall into two broad categories – those which fit an epidemic curve to past data in order to make predictions about the future, and mechanistic models which simulate more general scenarios based on specific underlying assumptions on factors such as varying contact rates or vaccine effectiveness. Both model types aid in the public health response, either by providing early warnings, or revealing the most important causes of spread and possible countermeasures. Stimulated by a flood of interdisciplinary research during the COVID-19 pandemic, researchers have also explored new approaches to epidemic forecasting, including models which exploit machine learning.
In a recent study, LML Fellow Colm Connaughton, working with Kirstin Roster and Francisco Rodrigues of the University of Säo Paulo, consider the role of so-called transfer learning for pandemic modelling. This refers to a set of techniques which apply knowledge from one prediction problem to solve another. For example, a model trained to execute a particular task in one domain can be used to perform a different task in another domain. The underlying idea is that skills developed in one task – learning the features relevant to recognizing human faces in images – may also be useful in other situations, such as classification of emotions from facial expressions.
As the researchers note, only a few studies so far have explored the potential of transfer learning for infectious disease modelling. In one study, for example, researchers trained a neural network on dengue fever time series and then made forecasts directly for two other mosquito-borne diseases, Zika and Chikungunya, in two Brazilian cities. Even without any data on the two target diseases, their model achieved high prediction accuracy four weeks ahead. This and other related studies suggest that transfer learning could be a valuable tool for epidemic forecasting in low-data situations.
In their recent study, Connaughton and colleagues contribute to this empirical literature by comparing different types of knowledge transfer and forecasting algorithms, and also considering two different pairs of endemic and novel diseases observed in Brazilian cities, specifically (i) dengue and Zika, and (ii) influenza and COVID-19. Overall, their results suggest that transfer learning methods show significant potential to improve disease forecasts, especially in the early phases of an epidemic when available data is limited. In addition, this approach may also offer a powerful means to adjust forecasts to new disease variants, such as the delta and omicron variants of Sars-COV-2.
The paper is available at

Leave a Reply

Your email address will not be published. Required fields are marked *