Influenza A/H3N2 epidemiology in England during the 2025 to 2026 season: a mathematical modelling study

Epidemiological analysesData summary

We analysed weekly counts of reported influenza cases by subtype for England from the Respiratory DataMart system from 2009-04-27 to 2025-12-08 (Fig. S1). We also analysed weekly counts of reported influenza specimens stratified by subtype for England from the WHO FluNet platform between 2011-01-09 and 2025-12-1413. All data are aggregated by week, and we used the first day of the epidemiological week as the reported date. We removed the most recent week of data to avoid issues arising from reporting delays. As many influenza A cases are not subtyped, we distributed unsubtyped counts into influenza A/H3N2 and A/H1N1pdm09 counts proportional to the ratios of subtyped samples.

To understand age-specific dynamics, we also used data on influenza cases by age group and by subtype from the Royal College of General Practitioners Research & Surveillance Centre (RCGP RSC) and ILI from the RCGP RSC weekly reports (Figs. S1–3). Following recommendations from12, we implemented an ILI+ indicator for each age group as:

$${{ILI}}_{+}={ILI}* {p}_{+{ve}}* N$$

(1)

Where \({p}_{+{ve}}\) is the proportion of all tests which are positive for influenza either overall or by subtype and \(N\) is the number of individuals in that age group. Note that imperfect case ascertainment is implicit in the ILI data.

Second-Generation Surveillance System (SGSS)

Weekly influenza positivity from PCR tests by age was obtained from the UKHSA dashboard for the SGSS, available at ref. 14. Values are shown as a percentage of people with at least one positive PCR result for influenza (in a 7-day period). Further details are available at ref. 15. This dataset provided us with the proportion of tests that are positive for influenza, but not the absolute number of tests performed, the proportion positive stratified by influenza subtype, nor stratification by age.

Royal College of General Practitioners (RCGP) Research & Surveillance Centre (RSC)

The RCGP RSC is a nationally representative primary care-based surveillance system. Weekly communicable and respiratory disease reports are publicly available at ref. 16. We compiled weekly national incidence of influenza-like illness (ILI) rates (per 100,000 population) for England. PDF reports were downloaded at 6-month intervals from January 2021 to December 2025, and Table E from each report was converted to CSV. These tables included age-stratified ILI rates across four age categories differing by reporting period:

1. (1)

   1–4 years, 5–14 years, 15–64 years and 65+ years, and all ages from week 26, 2023 to week 44, 2025.

2. (2)

   <15 years, 15–64 years, 65+ years, and all ages for weeks 40–53, 2020, week 1, 2021 to week 5, 2021, and weeks 1–26, 2023.

This dataset provides ILI rates by age, which we convert to estimates for absolute ILI case numbers by multiplying by the population size per age group17. To align different age bands from the two reporting periods, we redistributed the older ILI rates and percentage positive for influenza into the more recent age bands. First, we expanded age bands into per-year groups, assuming that the value reported within an age group applied to all ages within that group. We then recombined these per-year variables into common age bands (1–4, 5–14, 15–64 and 65+ years), either by summing counts and denominators within the new age band for incidence data, or by calculating the population-weighted mean using the per-year age distribution of England for positivity data17.

We also digitised data from the RCGP Virology Dashboard on the total number of samples tested and the number of samples positive for influenza by age group and by subtype for the 2022/2023 to 2025/2026 influenza seasons18.

Respiratory DataMart sentinel system for hospital testing

The Respiratory DataMart is a sentinel laboratory surveillance system which monitors all major respiratory viruses in England, compiling data from 17 contributing laboratories19. Participating laboratories test swabs for respiratory viruses, including influenza A viruses, using real-time polymerase chain reaction (RT-PCR), though not all laboratories test for or report all viruses. The absolute number of samples in this system is much lower than SGSS, but it gives absolute numbers of tests rather than just percentage positive.

We created a combined dataset spanning the period 27 April 2009 to 8th December 202514. This included weekly counts of laboratory-confirmed influenza cases by subtype: influenza A (not subtyped), influenza A(H1N1)pdm09, influenza A(H3N2), and influenza B. Additionally, the dataset provides the overall percentage of specimens testing positive for influenza. However, these data were not stratified by age group.

World Health Organisation (WHO) FluNet

Weekly counts of reported influenza specimens were obtained from the WHO FluNet platform13. This is the global influenza virological surveillance system collecting weekly sentinel and non-sentinel (e.g., outbreak investigations, point-of-care testing) surveillance data from national influenza centres. Data were extracted for the period 9th January 2011 to 14th December 2025, restricted to England. The extracted dataset included the weekly number of samples classified as A(H3), influenza A (not subtyped), A(H1N1pdm09) influenza B, and overall influenza cases. We allocated the unsubtyped influenza A cases into A/H3N2 and A/H1N1 cases proportional to the ratio of subtyped samples.

Weekly growth rate calculations

Empirical weekly growth rates were calculated for each influenza season and aligned by calendar week as:

$$y=log \left(\frac{i\left(t\right)}{i\left(t-1\right)}\right)$$

(2)

where \(i(t)\) is the reported incidence over week t. We calculated smoothed weekly exponential growth rates of influenza cases in England from the Respiratory DataMart using a Gaussian random walk model fitted to \(i(t)\) as described by Eales et al.20. Note that this model uses a prior distribution which penalises changes in the first derivative (approximately the growth rate). We also calculated overall weekly growth rates using data from the WHO FluNet for comparison using the same method. Finally, we fitted a Generalised Additive Model (GAM) to \(i(t)\) with a thin-plate spline using the mgcv R package to describe smoothed weekly growth rates of the age-stratified ILI+ data from the RCGP RSC data21.

Time-varying reproduction number estimation

We estimated the time-varying reproduction number (Rt) using the EpiEstim package in R22. Rt is defined as the expected number of new infections at time t, divided by past incidence weighted by relative infectiousness given by the generation interval distribution. EpiEstim uses an expectation-maximisation algorithm to estimate parameters of the following renewal equation given incidence data22,23:

$${I}_{t} \sim {Pois}\left({R}_{t}\mathop{\sum }\limits_{s=1}^{t}{I}_{t-s}{g}_{s}\right)$$

(3)

Where \({I}_{t}\) denotes the case incidence at time t; \({I}_{t-s}\) is the past incidence; \(\mathrm{Pois}\) refers to the Poisson distribution; and \({g}_{s}\) is the probability mass of the serial interval distribution at time-since-infection s.

We used the overall weekly influenza incidence data for the period 2011/2012 season to the 2025/2026 season from the UKHSA Respiratory DataMart for this analysis. To approximate daily incidence, weekly counts were distributed evenly across days per week and then smoothed with a 14-day rolling mean. The serial interval distribution was assumed to follow a distribution with a mean of 3.6 days and a standard deviation of 1.6 days, based on ref. 24. We used a window size of 14 days and the default Gamma-distributed prior on Rt with a mean of 5 and standard deviation of 522.

Compartmental model and scenario analysesModel overview

To explore potential epidemic scenarios distinguishing the current influenza season from the previous A/H3N2 season in 2022/2023, we simulated seasonal influenza transmission dynamics for England using an age- and immunity-structured deterministic Susceptible-Infected-Recovered model. We explored scenarios where the K subclade viruses exhibit enhanced transmissibility, greater immune escape, or an earlier epidemic seed time. We compared the impact of these scenarios on the epidemic timing around the half-term and Christmas school holidays, the final epidemic size, the final size in 65+ year olds, and peak incidence as a proxy for maximum healthcare burden, assuming that peak infection incidence is proportional to peak hospitalisation rate. We emphasise that we did attempt to formally compare model-predicted growth rates to observed growth rates based on estimates from the “Epidemiological analysis” section.

Model structure

We divided the population into four age groups (0–4, 5–18, 19–64 and 65+ years) and two immunity classes (fully susceptible and partially immune), representing eight classes in total.

The force of infection in group i was defined as:

$${\lambda }_{i}\left(t\right)=\beta \mathop{\sum }\limits_{j=1}^{m}{C}_{i,j}\left(t\right){I}_{j}\left(t\right)$$

(4)

Where β is the overall transmission rate (not age-stratified), Ci,j is the contact rate between group i and group j, and Ij is the number of infected individuals in group j at time t. The basic reproduction number, R0, was given by the largest eigenvalue of the next-generation matrix with contact matrix C, multiplied by β and the infectious period, Tg. The model assumes random mixing between individuals (mass action).

Transition rates between the three compartments were defined by the following set of ordinary differential equations:

$$\frac{d{S}_{a,k}}{{dt}}=-{\eta }_{k}{S}_{a,k}\left(t\right){\lambda }_{a,k}\left(t\right)$$

(5)

$$\frac{d{I}_{a,k}}{{dt}}={\eta }_{k}{S}_{a,k}\left(t\right){\lambda }_{a,k}\left(t\right)-\frac{{I}_{a,k}\left(t\right)}{{T}_{g}}$$

(6)

$$\frac{d{R}_{a,k}}{{dt}}=\frac{{I}_{a,k}\left(t\right)}{{T}_{g}}$$

(7)

Where \({\eta }_{k}\) denotes the relative susceptibility of immune class k and Tg is the infectious period. \({\eta }_{k}\) is a scaling factor which modulates the force of infection, \({\lambda }_{a,k}\left(t\right)\), for age group a and immune class k, where values less than 1 represent partial protection and values greater than 1 represent enhanced susceptibility. We modelled two immunity classes, \(k\in \{\mathrm{1,2}\}\) where \(k=1\) represents the susceptible class and \(k=2\) represents the immune class. We set \({\eta }_{1}\) to 1 for the susceptible population and \({\eta }_{2}\) to 0 for the immune population, representing all-or-nothing immunity (i.e., all immune individuals at baseline are unable to become infected). Note that \({\eta }_{k}\) is therefore effectively not used in the model but is included for completeness and to allow extension to scenarios assuming leaky immunity.

Overall immune escape was modelled as a single parameter, δ, which scales the initial population immune proportion in all age groups (δ = 0 corresponds to complete immune escape, whereas δ = 1 corresponds to no loss of population immunity relative to baseline). This parameter is used to scale the initial model conditions as:

$${N}_{a,1}\left(0\right)=\left(1-\delta \right){N}_{a},{N}_{a,2}\left(0\right)={\delta N}_{a}$$

(8)

\({N}_{a,k}\left(0\right)\) is then split across \({S}_{a,k}\left(0\right)\), \({I}_{a,k}\left(0\right)\), and \({R}_{a,k}\left(0\right)\) based on the proportions shown in Table S1. We solved the model in daily timesteps using the deSolve R package25.

Baseline model calibration

We used our intuition and manual calibration to generate a baseline scenario with epidemiological dynamics similar to what was seen in the 2022/2023 season. As a starting point, initial model parameters were based on standard seasonal influenza parameter values (R0 of around 2, infectious period of 4–5 days, and final size of around 15%26,27). We then manually adjusted the model parameters to broadly align the model-predicted influenza case incidence curves with the age-stratified ILI+ data from the 2022/2023 season based on four criteria: (1) the date of peak reported cases; (2) the magnitude of peak reported cases; (3) the cumulative number of reported cases; and (4) ensuring that the overall final size was between 15% and 20%. We also calculated the root mean squared error (RMSE) and RMSE normalised by the peak in reported cases between the model-predicted incidence curves and observed data. We note that this is a complex model with a large number of highly correlated parameters, making formal model fitting difficult. Many of the parameters are hard to identify and interpret, such as the overall fraction of symptomatic cases reported, the symptomatic fraction by age combined with age-specific reporting rates, and the level of immune escape of the seed virus. Thus, we focused on comparing outputs such as the timing and relative magnitude of peak reported cases and the overall proportion infected to the baseline 2022/2023 scenario, rather than the growth rate or precise parameter inference, by varying a limited number of key parameters. Parameter values used for the baseline scenario are shown in Table S1.

We set the population size of the model to 60,000,000 to approximate the population size of England. We distributed the population into age groups based on the age distributions in the socialmixr package using the POLYMOD data28. Each age group was stratified into the susceptible or fully immune class, assuming different levels of immune escape for each age group. The epidemic was seeded by arbitrarily setting I19-64(0) = 1000; the seed time was varied during model calibration.

Contact matrices over time

Symmetric, age-stratified contact matrices were generated using all contact data from the POLYMOD UK study using the socialmixr R package28. We generated four contact matrices for different time periods: (1) regular school term-time; (2) half-term with no school contacts and reduced school contacts; (3) pre-Christmas shopping period (2022-12-01 to 2022-12-21) with an increase in all non-school contacts; and (4) the Christmas school holiday period (2022-12-21 to 2023-01-05) with no school contacts, a reduction in all contacts, and an increase in at-home contacts. These matrices were constructed by resampling the original POLYMOD contact diary entries with replacement and applying multipliers for home, work and other contact types. We smoothed the transition between contact matrices over 3 days before and after the holiday period using a cosine function. Holiday dates were based on the Oxfordshire school holiday period. For all non-school contacts, the relative changes in contacts over half-term holidays and the Christmas period (shopping and holiday) were extrapolated from 2021–2023 data from Kendall et al.29. Absolute changes in overall non-school contact rates were extrapolated from the same source.

Scenario analyses

Scenario analyses were chosen to illustrate potential hypotheses for the early and rapid growth of A/H3N2 cases in England for the 2025/2026 season. We also varied the immune escape scaling parameter δ (testing all values from 0 to 1 in increments of 0.01), the basic reproduction number R0 (testing all values from 1.01 to 3 in increments of 0.01), the proportion of the 0–4 and 5–18 year old population initially immune (testing all values from 0 to 1 in increments of 0.01), and the seed date univariably (testing seed dates from 2022-08-01 to 2022-11-01 in increments of 1 day).

Use of artificial intelligence

Artificial intelligence (AI) tools including ChatGPT (GPT-5.1) and GitHub Copilot plugin for RStudio (GPT-4 and GPT-4o) were used to assist with literature review, to identify potential data sources and code development. All Large Language Model (LLM) output was checked and validated manually and modified by the authors where necessary. These tools were not used to generate data or interpret the findings, and no AI-generated content was used without human verification.