Functional requirement analysis report of Atrial Fibrillation Use Case

Atrial fibrillation diagnosis, affects more than 14 million over-65s for which the European Society of Cardiology has risen need for urgent action. As computational power increases, machine learning techniques are becoming more and more present and are integrated into many fields, especially in the healthcare sector.  Machine learning has been used previously for both risk prediction and identification of the phenotypes of atrial fibrillation (AF). However, the resulting machine learning models were not externally validated or showed moderate predictive ability and high risk of bias in an external validation. Undoubtedly, there is room for improvement for the AF. We propose an unsupervised machine learning technique applied to an integrative model including clinical data, imaging data, electrocardiogram (ECG) signals and genetic variants. We will detect phenotypes within the population of AF using an initial cohort in a first stage, and then, extending the study to other cohorts to generalize the model in a federated learning scheme. In the context of the HealthyCloud, we will show the whole process from the data discoverability to the technical part and the development of the machine learning models. This will have a high impact in the other WPs, particularly in WP2 and WP5, considering a real case going through all the steps of implementation. With the inclusion of the functional requirements of the AF use case mapped as analysis requirements, WP5 can have a better understanding in order to perform a broad analysis of existing and planned computational solutions, in terms of both infrastructures for research and advanced data analysis. The legal barriers, we had to face to implement a federated learning, will be a potential source of information for WP2 that will incorporate by design the ethical and legal considerations. The discoverability of data is also a key point for other WPs which focus on how data is structured, organised, and accessed either individually (WP3), or through data hubs (WP4) and/or potentially discovered through the FAIR health data portal (WP6).