An adaptive weight ensemble approach to forecast influenza activity in the context of irregular seasonality

Abstract

Forecasting of influenza activity in tropical and subtropical regions such as Hong Kong is challenging due to irregular seasonality with high variability in the onset of influenza epidemics, and potential summer activity. To overcome this challenge, we develop a diverse set of statistical, machine learning and deep learning approaches to forecast influenza activity in Hong Kong 0-to 8- week ahead, leveraging a unique multi-year surveillance record spanning 34 winter and summer epidemics from 1998-2019. We develop a simple average ensemble (SAE), which is the average of individual forecasts from the top three individual models. We also consider an adaptive weight blending ensemble (AWBE) that allows for dynamic updates of each model contribution based on LASSO regression and uses decaying weights in historical data to capture rapid change in influenza activity. Overall, across all 9 weeks of horizon, all models outperform the baseline constant incidence model, reducing the root mean square error (RMSE) by 23%-29% and weighted interval score (WIS) by 25%-31%. The SAE ensemble only slightly better than individual models, reducing RMSE and WIS by 29%. The AWBE ensemble reduce RMSE by 45% and WIS by 46%, and outperform individual models for forecasts of epidemic trends (growing, flat, descending), and during both winter and summer seasons. Using the post-COVID surveillance data in 2023-2024 as another test period, the AWBE ensemble still reduces RMSE by 32% and WIS by 36%. Our framework contributes to the ensemble forecasting of infectious diseases with irregular seasonality.