We conducted a retrospective cohort study using Scottish national laboratory and hospital registry data from 1 July 2017 to 30 June 2023.

Eligible persons in Scotland were offered the RSV vaccine from 1 August 2024 [6]. Prior to this, no vaccines against hMPV or RSV were implemented through the national immunisation programme during the study period.

We defined adults as individuals aged ≥ 18 years.

Comorbidity status was defined as those in clinical risk groups eligible for influenza vaccination according to the UK Green Book guidelines [10]. This includes individuals with specified chronic conditions such as cardiovascular, respiratory, renal, hepatic, neurological disease, diabetes, immunosuppression, and morbid obesity.

We included inpatient and non-routine hospitalisation episodes. Day cases and routine admissions were excluded from the analysis, using the relevant classification codes, as these are expected to have different resource utilisation patterns and a low likelihood of representing hospitalisation linked to RSV or hMPV infection.

Virus-specific hospital episodes were defined as admissions with a positive RSV or hMPV test from a specimen collected within 7 days before admission or up to 3 days after admission. The 7-day pre-admission window was chosen to reflect the known incubation period of RSV and hMPV (approximately 3–6 days), allowing infections acquired shortly before admission to be attributed to the hospital episode [11]. The 3-day post-admission window was included to capture patients tested shortly after admission, while minimising misclassification of hospital-acquired (nosocomial) infections. A 28-day washout period was applied at the patient level so that repeat laboratory tests within 28 days were considered part of the same infection; only positive tests separated by more than 28 days were counted as distinct test results.

As with our previous analyses, we defined an (annual) season as the period between 1 July of one year and 30 June of the next year [12]. The first three seasons (2017/18, 2018/19, and 2019/20) were classified as pre-COVID-19 pandemic seasons, and the subsequent three seasons (2020/21, 2021/22, and 2022/23) as COVID-19 pandemic/post-COVID-19 pandemic seasons, long length of stay (LOS) was defined as a hospital stay of more than 5 days [13].

Intensive care unit (ICU) admissions included both ICU and high-dependency unit (HDU) admissions within RSV and hMPV hospitalisation episodes.

Ninety-day mortality was defined as the deaths occurring up to 90 days post-discharge, including in-hospital deaths. Stratification by in-hospital and post-discharge deaths was not feasible in the analysis because of low counts (< 5) in multiple groups, leading to unstable estimates and data protection concerns.

Hospital episode data were obtained from the Scottish Morbidity Records 01 (SMR01), which includes all inpatient and day case episodes from non-obstetric and non-psychiatric specialties in Scotland [14].

Laboratory-confirmed RSV and hMPV infections were obtained from the Electronic Communication of Surveillance Scotland (ECOSS), based on routine clinical testing [15].

ICU or HDU admission data were obtained from the Scottish Intensive Care Society Audit Group (SICSAG) [16]. Mortality data were obtained from the National Records of Scotland (NRS) [17].

Data were linked using pseudonymised patient identifiers. SMR01 and ECOSS records were additionally matched using admission dates and specimen collection dates (as described above). SMR01 (hospital admission episode) records were linked to SICSAG (ICU/HDU) data using hospital admission dates and corresponding ICU/HDU admission dates, within the same hospital stay.

The Systematised Nomenclature of Medicine—Clinical Terms (SNOMED CT) codes for individuals eligible for influenza vaccination during the 2025 winter season were obtained from the NHS Digital Cohorting as a Service (CaaS) platform [18]. Although defined for England, these criteria align with the Joint Committee on Vaccination and Immunisation (JCVI) recommendations and apply to Scotland [10, 19].

SNOMED CT concepts were mapped to ICD-10-CM codes using the SNOMED CT Snapshot Extended Map Reference Set and linked using SNOMED CT identifiers [20]. Where multiple mappings existed, the highest-priority mapping was selected based on the Map_Priority variable.

Population estimates used to calculate incidence rate ratios (IRR) were obtained from the open-access National Records of Scotland [21, 22]. For each season, mid-year population figures corresponding to the end of the season were used.

The Public Benefit and Privacy Panel for Health and Social Care (PBPP-HSC) approved data access to the required data in the National Safe Haven (Approval number 2223-0019). Data were linked using pseudonymised patient identifiers assigned before data receipt. The active consent of patients was not required. The data were collected routinely through NHS Scotland, with patients informed of potential use and their rights through various PHS and NHS Scotland privacy notices.

All data processing and analyses were conducted using RStudio (version 4.5.3).

A complete-case approach was used, whereby observations with missing data for any variables included in a given model were excluded from that analysis.

We used negative binomial regression to estimate adjusted relative rates of hospitalisation by virus, age, number of comorbidities (0, 1, > 1), and season. Because population denominators stratified jointly by age and comorbidity were unavailable, age-specific population denominators were used as exposure offsets. Consequently, estimates for comorbidity represent adjusted relative hospitalisation rates based on the available season and age-specific denominators and should not be interpreted as true population incidence rate ratios by comorbidity. Profile likelihood confidence intervals were used to improve robustness in the presence of sparse data and small cell counts.

Observed proportions for intensive care unit (ICU) admission, prolonged hospital length of stay (LOS), and 90-day mortality were calculated as the number of events divided by the total number of hospital episodes within each stratum defined by virus type, age group, and comorbidity count. Low event counts did not allow further stratification by season leading to data protection concerns and unstable estimates. Estimates were expressed as percentages for interpretability. The 95% CI were estimated using nonparametric bootstrap resampling.

Logistic regression was used to examine associations between virus type and severity outcomes, adjusting for age, sex, SIMD, comorbidity count, and pandemic period. Ethnicity and season were excluded because of sparse data. Odds ratios and 95% confidence intervals were estimated using profile likelihood methods to improve robustness under sparse data conditions. Model performance was assessed using area under the receiver operating characteristic curve (AUC), Hosmer–Lemeshow tests, generalized variance inflation factor (GVIF), and diagnostics for separation and influential observations [23,24,25].

Inference was based on effect sizes and 95% CI. P values were not reported, as emphasis was placed on estimation rather than formal hypothesis testing.