Protocol: Federated analysis of Randomized Controlled Trial and Real-World data across two Trusted Research Environments to compare comorbidities in older adults vaccinated against COVID-19

Title: Federated analysis of Randomized Controlled Trial and RealWorld data across two Trusted Research Environments to compare comorbidities in older adults vaccinated against COVID-19 Authors: C. Ohmann, M. Panagiotopoulou, S. Contrino, L. Cudilla (ECRIN) C. Hall, M. Guignard-Duff, C. Cole (University of Dundee) P. Verde (University of Düsseldorf) Date: 15 July 2025 Version: 1 1. Background The EOSC-ENTRUST project 1 aims to create a European network of Trusted Research Environments (TREs) for sensitive data and drive European interoperability through the development of a common blueprint for federated data access and analysis. EOSC-ENTRUST has four drivers, or use cases, which are prototypic for federated, multinational use of TREs in research practice across scientific domains and user communities. Driver 3 demonstrates the potential ability of the blueprint to bridge the traditionally very separated data domains of clinical trials and real-world health data in one solution architecture. The proposed study serves as a prototype for EOSC-ENTRUST Driver 3, utilising individual participant data (IPD) from a randomized controlled trial (RCT) hosted in crDSR (TSD TRE), along with real-world data (RWD) hosted in the Health Informatics Centre (HIC), University of Dundee TRE with the aim of comparing comorbidities in older adults vaccinated against COVID-19. In the context of this protocol, we use the word “federation” in its broadest sense of connecting organisations together under a set of rules and standards that enable the analysis in the Driver 3 prototype. 2. Study objectives Overall Aim: To explore and evaluate different analytical approaches for federating two existing TREs in a real-world use case, addressing a clinically relevant question. Specific Objectives: 1 https://eosc-entrust.eu/ a) Investigate analytical approaches for federating two TREs, operating in two different countries (TSD 2 in Norway and HIC 3 in Scotland), for an analysis involving two different data types (RCT and RWD). b) Perform an evidence synthesis comparing IPD from a RCT with RWD filtered by similar inclusion criteria to assess the generalizability of the trial results on comorbidities in elderly individuals vaccinated against COVID-19. 3. Study design This study utilises existing datasets from a completed RCT and from routine healthcare vaccination records (RWD). 4. Study population Participants meeting the following criteria will be included: ● ≥ 75 years old ● Received 4 vaccination doses against COVID-19 For the RCT, these criteria define the inclusion; for the RWD, a filtering strategy will be applied to approximate the RCT cohort. 5. Datasets 5.1 Randomized Controlled Trial (RCT) ● Study: Assessing Immune Response of Different COVID-19 Vaccines in Older Adults (EU-COVAT-1) (NCT05160766) 4 ● Data controller: University of Cologne ● Data source: crDSR 5 (TSD TRE in Norway) ● Population: 269 participants ● Metadata: Documented in crDSR under 6 2 https://www.uio.no/english/services/it/research/sensitive-data/ 3 https://www.dundee.ac.uk/hic 4 Neuhann JM, Stemler J, Carcas AJ, Frías-Iniesta J. et al. Immunogenicity and reactogenicity of a first booster with BNT162b2 or full-dose mRNA-1273: A randomised VACCELERATE trial in adults ≥75 years (EU-COVAT-1). Vaccine. 2023;41(48):7166-7175. doi: 10.1016/j.vaccine.2023.10.029. 5 https://crdsr.ecrin.org/login 6 https://crdsr.ecrin.org/browsing/studies/DSRS-2/view ● Data de-identification status: Anonymised ● Key inclusion criteria: - Age ≥ 75 years. - Priming with homologous ChAdOx-1-S, BNT162b2, or mRNA-1273. - First booster with either BNT162b2 or mRNA-1273 at least 1 month prior to enrolment. - No SARS-CoV-2 infection within the previous 3 months. ● Key exclusion criteria: - Severe immunosuppressive therapy (e.g., high-dose glucocorticosteroids, active cancer treatment) 5.2 Real-World Data (RWD) ● Data controller: National Health Service (NHS) Tayside, Scotland ● Data source: Vaccination records of the Tayside population stored in HIC TRE in Scotland ● Population: 411,643 individuals ● Metadata: Documented in Healthdatagateway 7 ● Data de-identification status: Anonymised 6. Data to be analysed 6.1 Baseline data Baseline data will include basic variables such as sex, height, weight, body mass index (BMI), hospital admissions, prescribed medications, and laboratory results. The detailed list of variables to be compared will be finalized based on availability after feedback from the crDSR and the HIC teams. The GDPR data minimisation principle (Article 5(1)(c)) will be respected for the final selection of the relevant variables. 6.2 Vaccinations In the RCT, participants were randomized to receive a 4th dose dose of either BNT162b2 (Comirnaty, 30 μg) or mRNA-1273 (Spikevax, 100 μg). In the RWD, no restrictions are applied regarding the type of booster vaccination received, but subgroup analyses based on the different vaccine types are planned. The GDPR data minimisation principle (Article 5(1)(c)) will be respected for the final selection of the relevant variables. 7 https://healthdatagateway.org/en/data-custodian/10 6.3 Outcome criterion Between the RCT and the RWD, only one common outcome criterion, comorbidity 8 , is used to compare different TRE analytical approaches and to assess the generalizability of the trial results to the population in Tayside. Comorbidity includes, for example, cardiac disorders, metabolism and nutrition disorders, and vascular disorders. 7. Analytical models to be investigated In IPD meta-analysis two main analytical models exist 9 : - 2-stage: the IPD are first analysed separately within each study to obtain aggregate data (e.g., treatment effect estimates and standard errors); then, in the second stage, these aggregate data are combined in a standard meta ‐ analysis model (e.g., common ‐ effect or random ‐ effects). - 1-stage: the IPD from all studies are analysed in a single step using an appropriate model that accounts for clustering of participants within studies and, potentially, between ‐ study heterogeneity (e.g., a generalised linear mixed model). The implications of the different models are subjects of debate in the literature 10 . EOSC-ENTRUST organised a series of expert meetings and workshops to evaluate different possibilities for the analysis of IPD from clinical trials and routine healthcare, when these are stored in different TREs, located in different countries. On 15 January 2025, the “Support of IPD meta-analyses through Trusted Research Environments (TREs)” 11 took place. It was followed by a meeting in Utrecht of all the EOSCENTRUST Drivers. The conclusion of the meetings was that in practice, a framework for enabling research across TREs would either involve moving analyses to distributed datasets (“federated analytics”) or moving datasets into a single location for analysis (“data pooling”). The "lessons learned” of this pilot study seeking to analyse baseline characteristics, vaccination and comorbidity data held by two TREs will feed into the process of 8 HIC SMR01 - hospitalisation data for conditions encoded with ICD-10 9 Burke DL, Ensor J, Riley RD. Meta-analysis using individual participant data: one-stage and two-stage approaches, and why they may differ. Stat Med. 2017;36(5):855-875. doi: 10.1002/sim.7141. 10 Riley RD, Ensor J, Hattle M, Papadimitropoulou K, Morris TP. Two-stage or not two-stage? That is the question for IPD meta-analysis projects. Res Synth Methods. 2023;14(6):903-910. doi: 10.1002/jrsm.1661. 11 https://drive.google.com/file/d/1DdgQ7w61IB3UWtDMkvOkMo5qKBGGWPE2/view?usp=sharing defining requirements for TRE federation in EOSC-ENTRUST. This will be done by evaluating the feasibility and applicability of the analytical models below. The authors of this protocol are aware that applying in practice some of the models will be challenging (or even impossible) but an important outcome of our prototype involves documenting when a specific model cannot be applied, what the reasons are and which changes may be needed to enable analysis. Model no Analysis type Description Our prototype 1 2-stage Datasets will be analysed in the TRE holding them. The aggregated (and anonymous) results of the individual studies are collected and combined outside of the TREs. IPD analysed separately in the TSD TRE and the HIC TRE. Aggregated results combined outside of the TREs. Disclosure risk mitigation will be taken into consideration. 2 Data pooling Aggregate results from individual studies (analyzed in TREs) are pooled in a TRE within a federated network. IPD analysed separately in the TSD TRE and the HIC TRE. Aggregated results combined in one of the two TREs. Disclosure risk mitigation will be taken into consideration. 3 Remote execution (simple) A single TRE provides a view of IPD from multiple studies in the federated network, allowing the researcher to view and interact with the data. A virtual table is created from different sources, but full analysis will depend on the capabilities of the presenting tool. It’s up to the hosting TRE (and project governance) TSD and HIC are not yet part of a federated TRE network. In a theoretical exercise, to decide what the researcher can and can’t do. Model 3 is a simpler version of Model 4, and it is useful when remote queries can be easily encapsulated in a single message and datasets are homogeneous and tabular. requirements for Model 3 to be achieved will be collected. 4 Remote execution (generic) The analysis is executed in one TRE, but it accesses IPD stored in multiple other TREs within the federated network. More advanced than Model 3, allowing complex multi-study analyses without direct data movement. The main difference between Model 3 and Model 4 comes down to the complexity of the query sent between TREs. Model 3 works with simple queries that can be directly sent from TRE A to TRE B, while Model 4 handles more complex queries that can’t be packaged as a simple message. Instead, TRE A asks TRE B to run a specific workflow provided by a thirdparty service (e.g. Software Service C). Thus Model 4 adds an extra layer by involving a third party to handle complex workflows, while Model 3 relies on straightforward, direct queries. TSD and HIC are not yet part of a federated TRE network. In a theoretical exercise, requirements for Model 4 to be achieved will be collected. 5 1-stage IPD of all data sources are available and accessible in one TRE. The analysis will be performed in this TRE. Data transfer(s) are needed. In a theoretical exercise, requirements for Model 5 to be achieved will be collected. 8. Data Quality Assurance The RCT data have been monitored according to the Clinical Trial Regulation (CTR) 12 , including Good Clinical Practice (GCP) 13 . For the Tayside data, a “clean” 12 https://eur-lex.europa.eu/eli/reg/2014/536/2022-12-05 13 https://www.ema.europa.eu/en/ich-e6-good-clinical-practice-scientific-guideline and de-identified dataset will be provided according to the HIC internal procedures. 9. Statistical Methods Univariate and multivariate comparison of comorbidity as well as other variables (e.g., sex, age, height, weight, type of vaccination) will be performed. In order to be able to do this, a mapping procedure between different coding systems for diagnoses has to be performed. In the RCT dataset diagnoses are coded with the MedDRA system and in the RWD dataset with ICD-10. ICD-10 and MedDRA are two distinct medical terminologies used for different purposes, but they can be mapped to each other. The WHO 14 has developed a mapping between ICD-10 and MedDRA. It enables the integration and analysis of data from different sources using different terminologies. The mapping is available for download from the WHO and MedDRA 15 websites and will be used for the study. For univariate analysis the usual statistical tests will be performed: Chi-square test for a discrete variable, a t-test for continuous variable and a Kaplan-Meier test for time-toevent variable. In addition, a multivariate comparison between the n-dimensional distributions of the variables characterising comorbidity in both datasets will be performed using R. 10. Study timelines The present study will be performed according to the following steps: ● August 2025: Finalisation of the present study protocol with input from ECRIN, UNIVDUN and UiO. ● August-November 2025: Approval of the protocol by the responsible ethic committees and data protection authorities (if applicable according to the crDSR and HIC policies). Registration of the study protocol (e.g., Zenodo, OSF). ● August-December 2025: Data access request in the TREs (crDSR, HIC). ● December 2025-March 2026: IPD analysis to assess the generalisability of the RCT results and evaluate feasibility of the different TRE analytical models. ● March-May 2026: Summary of the results in a report and input of the study results into the requirements elucidation process of EOSC-ENTRUST. The exact timelines are subject to change depending on the data access approval process of each TRE involved. 14 https://cdn.who.int/media/docs/default-source/classification/icd/icd-10/icd-10-to-meddra-mapping-conventions.pdf?sfvrsn=1b36c13b_1 15 https://alt.meddra.org/files_acrobat/whatsnew_26_1_German.pdf 11. Ethics / Data protection As part of the study procedures, it will be checked whether approval of the study protocol by the responsible research ethics committees and data protection authorities is needed. If this is the case, the necessary approvals will be provided. It is expected that no informed consent from the individual participants is needed for the study but this also be checked in the study approval process. 12. Study team PI: S. Contrino (ECRIN), L. Cudilla (Access to TREs for analysis) Support staff: M. Panagiotopoulou (ECRIN) Advisors: C. Ohmann (ECRIN), P. Verde (University of Düsseldorf)