The tale of two assumptions: incorporating healthcare-seeking behaviour in epidemic forecasting

BACKGROUND: Modelling efforts during the COVID-19 pandemic highlighted the importance of incorporating human behaviour into mathematical models and the challenges of making accurate forecasts. Case detection is affected by different healthcare-seeking behaviors, including visiting physicians to seek help, which can impact the number of laboratory tests performed and the number of cases identified by surveillance systems throughout an epidemic. Mathematical models for forecasting epidemics of respiratory viruses such as influenza and COVID-19 generally assume a constant rate of case detection, and only a few studies have previously used time-dependent rates. PURPOSE: This study aims to compare constant and time-dependent case detection rate approaches for the forecast and retrospective fitting of seasonal influenza data. METHODS: An age-stratified Susceptible-Infected-Removed (SIR) model that incorporates case detection for influenza is formulated. Influenza case data and case detection rates for the 2016–2019 seasons in Alberta, Canada, are used for model training. The model fitting results are compared for the constant and time-dependent case detection assumptions. The model forecasting results using partial-season data are compared to the data for the remainder of the season for validation. RESULTS: While both constant and time-dependent case detection rate assumptions allowed an accurate retrospective fitting to the case data of an entire season, the forecasting performance showed a significant difference between the two assumptions. Models with a time-dependent case detection rate accurately predicted the influenza peak time four weeks before the actual peak occurred. The average total infections per case detected, an estimate that includes both under-ascertainment and underreporting, also showed a significant difference between the two assumptions. CONCLUSION: The incorporation of healthcare-seeking behaviour in mathematical modelling helps quantify the dynamic process of how infections are detected by surveillance systems. This is an important consideration for influenza forecasting. Since not all individuals engage in healthcare-seeking behaviour and only a fraction of those who do seek help may get tested, a proportion of infections remain undetected by the surveillance system. The retrospective forecasting results highlight that a time-dependent case detection rate is more representative of changes in healthcare-seeking behaviour during the influenza seasons than a constant case detection rate. This approach provides a reliable solution for improving forecasts of seasonal influenza and may be adaptable to other respiratory viral infections. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1186/s12879-025-11940-0.