Multi-stream data-driven forecasting of influenza activity and associated hospital admission burden: an implication for impact assessment of COVID-19 pandemic on 2019/20 winter influenza season in Hong Kong

Abstract

<p><strong>Introduction</strong></p><p>In theory, better forecasts or projection of influenza depends on understanding the impacts of the associated drivers and interventions, and their incorporation could empower the predictive performance of the underline models. Like other tropical and subtropical regions, influenza viruses can circulate year-round in Hong Kong. However, during the COVID-19 pandemic, there was a significant decrease in influenza activity.</p><p><br></p><p><strong>Objectives</strong></p><p>The objective of this study was to retrospectively forecast influenza activity during the year 2020 and assess the impact of COVID-19 public health social measures (PHSMs) on influenza activity and hospital admissions in Hong Kong.</p><p><br></p><p><strong>Methods:</strong></p><p>Using weekly surveillance data on influenza virus activity in Hong Kong from 2010 to 2019, we developed a statistical modeling framework to forecast influenza virus activity and associated hospital admissions. We conducted short-term forecasts (1-4 weeks ahead) and medium-term forecasts (1-13 weeks ahead) for the year 2020, assuming no PHSMs were implemented against COVID-19. We developed two frameworks based on (1) statistical modelling and (2) mechanistic modelling to construct the respective multiple streams data driven predictive models. We used the out-of-sample validation and temporal cross-validation techniques to check the forecasting performance. We estimated the reduction in transmissibility, peak magnitude, attack rates, and influenza-associated hospitalization rate resulting from these PHSMs.</p><p><br></p><p><strong>Results</strong></p><p>For short-term forecasts, mean ambient ozone concentration and school holidays were found to contribute to better prediction performance, while absolute humidity and ozone concentration improved the accuracy of medium-term forecasts. We found the influenza activity/transmissibility had the significant non-linear associations with the mean absolute humidity (U-shaped), mean ambient ozone concertation (negative power) in Hong Kong. We observed a maximum reduction of 44.6% (95% CI: 38.6% - 51.9%) in transmissibility, 75.5% (95% CI: 73.0% - 77.6%) in attack rate, 41.5% (95% CI: 13.9% - 55.7%) in peak magnitude, and 63.1% (95% CI: 59.3% - 66.3%) in cumulative influenza-associated hospitalizations during the winter-spring period of the 2019/2020 season in Hong Kong. We found the forecast outcomes in both frameworks are comparable with respective predictive accuracy.</p><p><br></p><p><strong>Conclusion:</strong></p><p>We developed the integrated frameworks to not only forecast influenza activity and hospitalizations but also project influenza activity and hospitalizations retrospectively under a counterfactual scenario without COVID-19 PHSMs since January 2020. The implementation of PHSMs to control COVID-19 had a substantial impact on influenza transmission and associated burden in Hong Kong.</p>