Research results

Description

Work Package 1 (WP1) is developing the methodology to involve citizens, professionals, patients, public authorities and other relevant actors in the evaluation and implementation of Artificial Intelligence (AI) tools for heart failure. The goal of this WP is to give voice to patients, healthcare professionals, caregivers, policy makers, researchers and engineers specializing in AI and have them discussing together what is needed to develop and implement a reliable and useful tool, capable of reaching out to people and improving their quality of life. The main outcomes of this WP are the first version of the information and communication package (T1.2) and the requirements for the development of trustworthy AI for heart failure (T1.3-T1.6).

Description

Work Package 2 (WP2) focuses on preparing and managing datasets from clinical sites while ensuring compliance with data policies such as FAIR (Findable, Accessible, Interoperable, and Reusable) and GDPR (General Data Protection Regulation). The WP aims to install and execute the federated learning (FL) platform across all technical and clinical sites throughout the project. Key activities within WP2 include data governance and management, data extraction and harmonization, and federated platform installation and execution. The University of Barcelona leads WP2, ensuring seamless data integration and harmonization among diverse clinical centers.

Description

Work package 3 (WP3) entails a cross-disciplinary effort between University of Oxford, University of Barcelona, and several other participating institutes within the AI4HF collaboration. This work package aims to employ a combination of federated learning and transfer learning on multimodal data from multiple AI4HF sites, to build trustworthy AI models that are usable across a diverse range of clinical settings for patients with heart failure (HF). Therefore, of critical importance to this WP is:

1. Development of multi-modal AI models designed for forecasting crucial clinical outcomes among HF patients.
2. Exploration of strategies to enhance generalizability, robustness, and fairness in these models. Crucial to achieving this goal, in order to enable adaptable deployment to diverse data sites, two approaches will be investigated: federated learning involving decentralized model training across data sites, and transfer learning where models trained in one setting are shifted to other data settings.
3. Development of techniques to quantify prediction uncertainty, thereby offering actionable confidence scores for end-users.

This WP3 will involve close collaboration with various clinical, ethics, UX, and other WP’s. Achievement of these objectives will lead to high-quality research in AI modelling for risk prediction in HF patients leading to novel models that have been tested thoroughly for various aspects of real-world deployment.

Description

Work Package 4 (WP4) specific objectives are: a) Implement AI4HF’s UX and UIs based on the requirements and trustworthy AI designs from WP1, taking into account the variety of needs, preferences and roles (clinicians vs. patients vs. data/IT managers), b) implement explainability methods, AI-patient interfaces and user-friendly manuals to maximise usability, acceptance, trust and adoption, and c) integrate all AI models from WP3 and interfaces from this WP into a Clinical Decision Support System (CDSS) for human-centred, AI-guided management of HF risks. In this direction, WP4 will implement user interfaces and visual analytics for clinician-AI and patient-AI interactions, to enhance informed shared decision making like never before, ensure wide and intuitive use of the tools, and increase patient literacy and adherence to care plans.

Description

Work package 5 aims to deliver the traceability and monitoring tools for post-deployment of AI models. Its first objective is to deliver the AI4HF passport for standardised description and increased transparency of the AI models. It will enable access to key information about the AI’s production and maintenance, including model properties and hyperparameters, training and testing datasets, evaluation metrics and results, biases and other limitations, ethical approvals and data governance. The second objective is to implement several monitoring functionalities for AI models, including data quality control, human oversight and feedback, continuous evaluation and learning, as well as time-series visual analytics. Finally, we will integrate all these modules into a user-friendly, ready-to-deploy AI4HF traceability suite for assessment and monitoring of the deployment and execution phase of AI models.

Description

Once the necessary AI4HF’ models and tools are developed, our focus is on the evaluation, promotion, and preparation of the AI4HF AI technology for real-world adoption. To that end, WP6 focuses on the implementation of a multi-center AI validation study, with unprecedented scale and depth. We will consider a variety of clinical settings (8 hospitals), subpopulations (multiple Heart Failure subgroups and aetiologies), datasets, ethnicities (European, South American, and African), and end-users (80 clinicians, 100 patients, 16 data/IT managers). Additionally, we will consider multiple criteria and dimensions of trustworthy AI, based on 6 main principles namely: Fairness, Universality, Traceability, Usability, Robustness, and Explainability. The proposed AI evaluation study will be structured based on four stages of increased complexity: (1) Feasibility, (2) Capability, (3) Applicability, and (4) Durability. At first, we will perform cross-validation studies with retrospective data to identify preliminary limitations and get feedback on how to enhance the first versions of the AI4HF tools, including the AI explainability methods, the user interfaces, and the AI monitoring tools. Second, we will assess the model’s capability under larger variability. Third, we will evaluate the tool’s applicability in clinical practice with end-users, by focusing on usability, explainability, acceptability, clinical integration, and clinical utility. Finally, we will evaluate the novel AI monitoring platform, in terms of the performance, utility and acceptance of its different functionalities (quality control, time-series visualizations, periodic evaluations, data drift detection). These stages will ensure the realistic implementation, reduce costs, gather continuous feedback and generate evidence incrementally on the strengths and limitations of the AI functionalities.

Description

Finally, having evaluated the AI tool clinically, empirical data and experiences from the evaluation studies will be used to assess various implications of the introduction of the AI4HF tools in future HF care, including cardiology costs, new clinical guidelines, socio-ethical aspects, as well as certification, exploitation and sustainability by:

• Cost evaluation for the AI4HF technology (Lead: ESC) by developing an ad hoc tool for cost assessment.
• Impact on clinical practice (Lead: Amsterdam UMC) by reviewing and mapping the current HF guidelines, implications of the AI4HF tools, a one-day workshop and whitepaper.
• Socio-ethical implications (Lead: SHINE) by analyzing the perceived transparency and trust in the technology, a one-day workshop and whitepaper.
• Plan for regulatory (CE) acceptance and approval (Lead: Regenold) by a CE-workshop and defining pathways.
• Delivering the final version of the information and communication package (Lead: Amsterdam UMC).
• Intellectual Property Rights (IPR) management, exploitation and sustainability planning (Lead: SRDC) by performing a market analysis of (existing) AI solutions and an IPR/Exploitation Working Group.

Description

Work package 8 aims to continuously disseminate and communicate the project and its results to clinical, research and industrial stakeholders, as well as to the wider public. Additionally, it is intended to create synergies with other initiatives and EU projects in the field of trustworthy medical AI. For these objectives, a strategy, and plan will be developed with corresponding materials. Moreover, WP8 also ensures to monitor the successful implementation of research activities (WP1-WP7) within the agreed time, costs and quality metrics, with continuous management of risks and corrective actions. Besides, all administrative, financial, and contractual aspects related to the project are coordinated and managed by WP8.