Periodic Reporting for period 2 - REDDIE (Real-world evidence for decisions in diabetes)
Período documentado: 2024-07-01 hasta 2025-12-31
Randomised controlled trials (RCT) are the cornerstone of evidence-based medicine. However, the digitisation of real-world data (RWD) including data from devices, wearables, and electronic health records in large national registries provides opportunities to demonstrate efficacy and safety of innovative technologies including drugs, devices, diagnostics, and digital health. These data are particularly relevant to long-term conditions such as diabetes mellitus, where drugs, lifestyle interventions, and digital technologies often work together.
To better utilise RWD in diabetes for regulatory decision making, a development of standards, guidance, and an assessment of the efficacy to effectiveness gap is needed.
REDDIE (Real-World Evidence for Decisions in Diabetes) aims to explore how RWD can complement RCTs to improve efficacy, safety, and value for money of technologies to prevent and treat diabetes. The overall aim of REDDIE is to support the use of RWD in diabetes and health-related research.
To better utilise RWD in diabetes for regulatory decision making, a development of standards, guidance, and an assessment of the efficacy to effectiveness gap is needed.
REDDIE (Real-World Evidence for Decisions in Diabetes) aims to explore how RWD can complement RCTs to improve efficacy, safety, and value for money of technologies to prevent and treat diabetes. The overall aim of REDDIE is to support the use of RWD in diabetes and health-related research.
Work performed and main achievements
The project achieved significant progress across several work packages:
• Mapping the Data Landscape: Researchers conducted a systematic review identifying 573 diabetes databases across 58 countries, primarily in Europe, North America, and Asia. An online dashboard was launched to make these resources discoverable to the global research community.
• Fit-for-Purpose Assessment: The team developed a structured three-stage tool (defining the research question, data extraction, and assessment) to evaluate whether a database contains the necessary variables and follow-up time to answer specific clinical questions.
• Regulatory and HTA Standards: REDDIE reviewed 14 existing guidance frameworks and identified the Akehurst et al. framework as a foundation for developing diabetes-specific standards. Stakeholder workshops were held to refine these into normative statements for future regulatory use.
• Trial Emulation: The project emulated landmark trials (such as LEADER, EMPA-REG, and DEVOTE) using routine clinical data from registries in Denmark, Sweden, the UK, and Germany. These emulations used longitudinal targeted maximum likelihood estimation (TMLE) to ensure robust causal inference.
Specifically, we have already published analysis plans and are presently working on an analysis of the following models in RWD:
1. GLP-1 vs. DPP-4 (LEADER inspired population; carried out in DK, UK, SE)
2. SGLT-2 vs. DPP-4 (EMPA-REG inspired population; carried out in DK, UK, SE)
3. Insulin degludec vs. glargine (DEVOTE inspired population; carried out in DK)
4. DPP-4 vs. SU (TECOS inspired population; carried out in DK)
5. GLP-1 vs. DPP-4 as second-line therapy (LEAD2 inspired population; carried out in DK, GE)
6. GLP-1 receptor agonists vs. metformin as first-line therapy (LEAD1 inspired population, carried out in DK, GE, SE)
Advanced Analytics: WP5 compared various confounding-adjustment strategies and developed an augmented Propensity Score Matching (PSM) approach that automates parameter tuning to improve scalability and reproducibility. Additionally, an R library was developed to build Dynamic Bayesian Network (DBN) models for simulating complex longitudinal dependencies.
The project achieved significant progress across several work packages:
• Mapping the Data Landscape: Researchers conducted a systematic review identifying 573 diabetes databases across 58 countries, primarily in Europe, North America, and Asia. An online dashboard was launched to make these resources discoverable to the global research community.
• Fit-for-Purpose Assessment: The team developed a structured three-stage tool (defining the research question, data extraction, and assessment) to evaluate whether a database contains the necessary variables and follow-up time to answer specific clinical questions.
• Regulatory and HTA Standards: REDDIE reviewed 14 existing guidance frameworks and identified the Akehurst et al. framework as a foundation for developing diabetes-specific standards. Stakeholder workshops were held to refine these into normative statements for future regulatory use.
• Trial Emulation: The project emulated landmark trials (such as LEADER, EMPA-REG, and DEVOTE) using routine clinical data from registries in Denmark, Sweden, the UK, and Germany. These emulations used longitudinal targeted maximum likelihood estimation (TMLE) to ensure robust causal inference.
Specifically, we have already published analysis plans and are presently working on an analysis of the following models in RWD:
1. GLP-1 vs. DPP-4 (LEADER inspired population; carried out in DK, UK, SE)
2. SGLT-2 vs. DPP-4 (EMPA-REG inspired population; carried out in DK, UK, SE)
3. Insulin degludec vs. glargine (DEVOTE inspired population; carried out in DK)
4. DPP-4 vs. SU (TECOS inspired population; carried out in DK)
5. GLP-1 vs. DPP-4 as second-line therapy (LEAD2 inspired population; carried out in DK, GE)
6. GLP-1 receptor agonists vs. metformin as first-line therapy (LEAD1 inspired population, carried out in DK, GE, SE)
Advanced Analytics: WP5 compared various confounding-adjustment strategies and developed an augmented Propensity Score Matching (PSM) approach that automates parameter tuning to improve scalability and reproducibility. Additionally, an R library was developed to build Dynamic Bayesian Network (DBN) models for simulating complex longitudinal dependencies.
Successfully developing and implementing the described new methods and evidentiary standards for RWD for regulatory and HTA decision-making has the potential to get new technologies to patients sooner and with fewer RCTs, which will have a significant effect on R&D cost.
• Empirical Validation of RWE: The trial emulations provided concrete evidence that RWD can strikingly reproduce RCT findings in high-risk populations. However, it also revealed that in "wider," lower-risk populations, treatment benefits were often attenuated or absent—a nuance that standard trials often miss.
• Synthetic Data Generation: By utilizing Dynamic Bayesian Networks, the project successfully created synthetic datasets that closely replicate the variable distributions and survival curves of original clinical trials, such as LEADER. This allows for data sharing and methodological testing without the privacy risks associated with real patient data.
Our development of modelling techniques using synthetic data as well as our simulation models based on machine learning will improve the conduction of retrospective observational studies based on RWD in scenarios where automation and scalability are important. Also, the enhanced bnstruct library developed in WP5 will provide researchers with an enhanced software to build DBN simulation models from RWD, facilitating the development of new simulation models and their use for in silico trials.
In conclusion, the REDDIE framework delivers a structured decision-making environment for chronic disease management that can be adapted beyond diabetes to other complex therapeutic areas.
• Empirical Validation of RWE: The trial emulations provided concrete evidence that RWD can strikingly reproduce RCT findings in high-risk populations. However, it also revealed that in "wider," lower-risk populations, treatment benefits were often attenuated or absent—a nuance that standard trials often miss.
• Synthetic Data Generation: By utilizing Dynamic Bayesian Networks, the project successfully created synthetic datasets that closely replicate the variable distributions and survival curves of original clinical trials, such as LEADER. This allows for data sharing and methodological testing without the privacy risks associated with real patient data.
Our development of modelling techniques using synthetic data as well as our simulation models based on machine learning will improve the conduction of retrospective observational studies based on RWD in scenarios where automation and scalability are important. Also, the enhanced bnstruct library developed in WP5 will provide researchers with an enhanced software to build DBN simulation models from RWD, facilitating the development of new simulation models and their use for in silico trials.
In conclusion, the REDDIE framework delivers a structured decision-making environment for chronic disease management that can be adapted beyond diabetes to other complex therapeutic areas.