• Q1 2024
  • Q3 2025
  • Q4 2025
  • Q1 2026
  • Q1 2026
  • Q3 2026
  • Q2 2027
  • Q4 2027
  • Q4 2028
  • >2028

dAIbetes start

Virtual twins may be used as prognostic tools in precision medicine for personalised disease management. However, their training is a data hungry endeavour requiring big data to be integrated across diverse sources, which in turn is hampered by privacy legislation such as the General Data Protection Regulation. Privacy-enhancing computational techniques, like federated learning, have recently emerged and hold the promise of enabling the effective use of big data while safeguarding sensitive patient information.

In dAIbetes, we build on this technology to develop a federated health data platform for clinical application of the first internationally trained federated virtual twin models. Our primary medical objective is personalised prediction of treatment outcomes in type 2 diabetes, which afflicts 1 in 10 adults worldwide and causes annual expenditures of ca. 893 billion EUR.

While healthcare providers are becoming increasingly effective at targeting diabetes risk factors (e.g. diet or exercises), no guidelines as to the expected outcome for a given treatment for a specific patient exist. To address this urgent, yet unmet need, the federated dAIbetes technology will harmonise existing data of ca. 800,000 type 2 diabetes patients of 4 cohorts distributed across the globe, and learn prognostic virtual twin models. Those will be validated for their clinical relevance and applied in clinical practice through a dedicated software.

We aim to demonstrate that our personalised predictions have a prediction error that is at least 10% lower than that of population average-based models. This federated virtual twin technology will enable personalised disease management and act as a blueprint for other complex diseases. Our consortium combines expertise in artificial intelligence, software development, privacy protection, cyber security, and diabetes and its treatment. Ultimately, we aim to resolve the antagonism of privacy and big data in cross-national diabetes research.

Guidelines for data subject information and consent

Ethics and data protection are at the centre of dAIbetes and the prime rationale behind its federated architecture.

Experts in human rights, ethics, security, privacy and law will be involved in the planning of all main activities of the project which have ethical, security, privacy or other human rights implications, in order to analyse the work in this project concomitantly from day 1 and devise recommendations, guidelines, risk analyses and mitigation measures, which will be continuously fed into the design and implementation processes of this project.

Data exchange apps

dAIbetes focus on the development of a federated database network (dAIbetes-Net) that will connect at least four isolated and globally distributed independent repositories with longitudinal data from type 2 diabetes patients.

One goal is the development of apps for standardised exchange of data. The output will be software applications (apps) that will be integrated at every participating local data repositories and hospitals of dAIbetes-Net, in order to facilitate import/export of data in a standardised format.

Clinical validation study

The overall clinical research objective of dAIbetes project is to develop the first highly integrated health data platform comprising clinical data registered in dAIbetes-Net, connected to federated dAIbetes learning apps and the first internationally unified and validated virtual twin models for personalised outcome prediction for type 2 diabetes treatment.

The dAIbetes software platform will be a tool for using the virtual twin models trained on the harmonised data of ca. 800,000 subjects for the prediction of primary (HbA1c) and secondary outcomes, which is to be used in clinical practice.

It then becomes important to assess prospectively, in an external validation, that the virtual twin-based predictive models operate as planned and are considered useful by clinicians in practice in real-world settings and environments. We will accomplish this with a validation/feasibility study for personalised outcome prediction after type 2 diabetes treatment of at least 3,600 patients.

Federated learning methods

dAIbetes focus on the development of virtual twin apps for training of virtual twin models that will use data from type 2 diabetes patients and a privacy-preserving training based on a federated database network.

We will benchmark and compare various federated learning methods based on the implemented dAIbetes virtual twin models. The models will be validated with optimal clinical (primary) outcome prediction as the (primary) target (and the secondary outcomes as secondary targets).

Both, the federated learning and the federated twin model learning apps will be made compatible with the EHDS, where possible.

Virtual twins implemented and tested

In a first step we will perform a comprehensive clinical outcome prediction evaluation, first for the primary and subsequently for all secondary outcomes. During the runtime of dAIbetes, we will also perform an external validation on prospective clinical data that the models have never seen for training.

Here, we aim to reach clinical relevance by showing that our personalised prognostic virtual twin models have a prediction error that is at least 10% lower than that of currently used population averages.

Robustness analysis

In dAIbetes we will study the robustness of the federated approaches against models trained in a centralised fashion, and in particular implement a location-wise robustness analysis (leave-one-hospital-out cross validation) to quantify the potential effect of the federated infrastructure regarding its susceptibility to distributed batch effects.

Open training and developer hackathon

A developer hackathon for young scientists will be organised to stimulate future app development.

Two workshops or webinars for clinicians and regulatory bodies will be offered at early stage of the project for awareness creation and a third one, at the end of the project, to educate the clinicians, pharmacists, researchers and interested GPs in the potential of the virtual twin models for personalised outcome prediction of type 2 diabetes treatment and for personalised therapy design.

Evaluation of clinical outcome prediction accuracy

We will first perform a comprehensive clinical outcome prediction evaluation, first for the primary and subsequently for all secondary outcomes. The evaluation will first be based on the existing (retrospective) data assets and labels, using internal cross-validation and bootstrapping schemes. The error function is the deviation of the predicted (quantitative) outcome from the observed outcome.

During the runtime of dAIbetes, we will also perform an external validation on prospective clinical data that the models have never seen for training.

beyond 2028

In dAIbetes, we leverage federated learning to build a federated health data platform, creating the first internationally trained virtual twin models for type 2 diabetes.

This innovative approach aims to personalize treatment outcome predictions, which currently lack precise guidelines.

© 2024 | All rights reserved
X   LinkedIn