Project

Challenge / SotA: In large extensive development projects, effective project management is a key factor contributing to successful delivery and implementation of results. The typical challenges include isolated working groups, lack of transparency in communications and late onset of result accumulation. The lack of information exchange within the consortium may generate duplicate efforts and suboptimal results, whereas sharing and publishing results at the very late stages of the projects leads to incomplete testing and validation. Having an integrated project workflow can prevent both of these consequences, which have direct ramifications on impact generation.

Proposition/ Ambition: The PHASE IV project management shall incorporate elements from the Ops-approach (DevOps, MLOps and DataOps). The work will be conducted in several development cycles and a consortiumwide evaluation of the results will be performed at the end of each cycle. These evaluation tasks correspond to the important project milestones. The integrated project workflow will be established leveraging the applicants’ experiences from previous projects and the relevant aspects, including internal communications, shall be described in the PHASE IV AI project handbook and PHASE IV AI exploitation plan.

Outcome: A validated transferable, integrated project workflow to support the execution of large development projects. This workflow will foster the development and implementation of the PHASE IV AI solutions and services according to user and stakeholder needs and requirements.

Challenge / SotA: Synthetic data can be used as a privacy-enhancing technology to provide individual-level data without compromising the privacy of data subjects. However, the process of generating synthetic datasets represents a well-known trade-off between utility and privacy. Generic synthetic data sets are only suited for a limited range of applications. What’s more, evaluating the quality of synthetic datasets is difficult, as there is no general consensus regarding the metrics to be applied for privacy and utility assurance. Especially when it comes to synthetic medical images, suitable metrics are non-existing.

Proposition/ Ambition: To avoid the problem of all-purpose synthetic datasets, PHASE IV AI will provide synthetic data on demand, which allows for optimizing the synthetic data quality for a pre-determined purpose. The confidentiality will be ensured using data synthesis methods that provide strong, verifiable privacy guarantees (e.g. differential privacy). PHASE IV AI will define a set of metrics to evaluate the quality of synthetic data, including the introduction of novel metrics especially for image data. Finally, the utility of synthetic datasets will be validated in 3 real-world use cases (Objective 5, WP6).

Outcome: Data as a Service (DaaS) concept and implementation.

Challenge / SotA: Researchers and health technology developers are struggling to overcome the disadvantages of small datasets such as low generalizability or statistical power. Combining datasets from different sources, however, is often impossible due to technical, operational, and legislative barriers. The baseline requirement is having inter-operable data to process and analyse. Secondly, deriving real-world evidence from distributed sources requires a secure connection between the sites, which can be achieved via secure multi-party computation (SMPC).
The current challenges in SMPC protocols are related to the maximum number of sites (nodes) and overall computational overhead. An additional challenge is linked to AI development and producing models (e.g. for synthetic data generation) that adequately mirror the statistical properties of all local datasets used for training.

Proposition/ Ambition: PHASE IV AI deploys and builds on an existing solution for secure multi-party computation (XOR Secret Computing® engine by INPH) as foundation of scaling-up SMPC. This already allows deriving insights from real-world data across multiple locations to privately join the data, compute aggregate statistics and/or build machine learning models. The industry partner will collaborate within the consortium’s expertise on processing specific healthcare data such as MRI images, clinical data as well as omics data. Within the consortium, INPH’s expertise can be combined with the expertise of the other consortium partners to generate synthetic data which will be essential for iterative development of the ML models. This requires work on two fronts: increasing the computational capacity via hardware acceleration and research on the methods for decentralized training of machine learning models (federated machine learning with secure aggregation by SMPC or FHE, or a combination of the two). To be able to evaluate the results, PHASE IV AI will also include research activities around model validation.

Outcome: Model as a Service (MaaS) concept and implementation of tools and innovative infrastructure for scalability.

Challenge / SotA: For new health technologies to enter the European market, health data is required at multiple steps of the process, from feasibility studies to health technology assessment. Being able to search for and access the relevant data in a privacy compliant format would save a lot of resources on the way, enabling a faster route from development to product launch. The barriers to data-driven innovation are highlighted in the case of small and medium-sized enterprises that rarely maintain extensive databases on their own.

Proposition/ Ambition: The PHASE IV AI services for privacy compliant health data, Data as a Service and Model as a Service, will be joined into a data market entity referred to as Health Data Hub. The core mission of the Health Data Hub is to support secondary use of health data by enhancing and simplifying data discovery and providing secure, on-demand access to anonymous as well as synthetic data as a service and paving the path for utilization of SMPC as a Model as a Service offering. Furthermore, the PHASE IV AI will foster data exchange since data will undergo a check using the PHASE IV AI quality metric indicating the trustworthy quality of the exchanged data. The easy utilization of these technical services will facilitate and the GDPR compliant secondary use of health data for research and innovation a lot.

Outcome: PHASE IV AI Services enabling easy to use GDPR-compliant high quality health data exchange.

Challenge / SotA: The utility of anonymous datasets, including synthetic data, must be demonstrated in practice. This is critical for revealing the potential technical issues, but also for assessing the societal readiness level. For example, synthetic data is only emerging as a tool for health technology developers and there are rightful concerns to be addressed. Testing and validation in real-world use cases will set realistic boundaries for the use of synthetic data as stand-alone training data or an extension to real-world data (augmentation).

Proposition/ Ambition: The PHASE IV AI use cases are designed to test and validate the utility of privacy compliant health data, including synthetic data and ML models produced through secure multi-party computation, in three high-impact diseases. Co-creation with the technology users such as AI developers and clinicians is an effective method to increase technology uptake. Acceptability studies of Data as a Service (WP3) and Model as a Service (WP4) will be conducted and the results of these studies will feed all stages from the design (WP2) to evaluation (WP6) for iterative improvement. PHASE IV AI will investigate initial user requirements for a diversity of users and collect feedback at the end of each development cycle. The selected clinical decision-support systems will be used with or without input from the envisioned PHASE IV AI health data services (DaaS and MaaS) and the comparisons serve as a basis for evaluation. The overall performance will also be compared with other stateof-the-art systems and published replication studies (synthetic data versus real-world data).

Outcome: Improved health technologies for diagnosis and treatment of the selected high-impact diseases. Continuous feedback & evaluation guiding development cycles in other work packages, in particular WP3, WP4.

Challenge / SotA: All PHASE IV AI activities should align with and complement the current European activities towards a common European Health Data Space. The large-scale impact can only be reached by engaging with the European health ecosystem and its key stakeholders. PHASE IV AI also recognizes the groundwork done in other projects such as TEHDAS, DARWIN and EHDEN. and will built on the existing foundations. For PHASE IV AI, the proposed AI Act represents an important development originating from outside the health industry. Active participation and communication will require inputs from all consortium partners.

Proposition/ Ambition: PHASE IV AI establishes and maintains an active communication with relevant stakeholders. The specific means for public outreach and stakeholder engagement will be defined in the PHASE IV AI dissemination plan. The PHASE IV AI Health Data Hub involves a Data Market for exchanging information about the data requirements and data availability. The co-creation approach adopted in the use cases (WP5, Objective 5) also serves end-user involvement and education on state-of-the-art technologies. The liaisons with external expert groups support this aim, including clinical stakeholders and expert groups.

Outcome: A user-friendly Health Data Hub facilitating easy to use privacy preserving exchange of high-quality health data enabling the innovation of data driven healthcare services.