Work Packages

Effective project management is a central element of successful research. This is because large research projects entail a lot of administrative work. The following objectives will be actively pursued by WP1:

• To make FeatureCloud achieve its objectives and to deliver in time, budget and quality its milestones and deliverables
• To help the consortium abide by the regulations and contractual obligations according to the grant agreement, its annexes and the consortium agreement
• To look after the project`s finances and to report them properly to the European Commission
• To establish a communication infrastructure which enables the partners to communicate efficiently and to stay connected for the run-time of the project
• To preserve the rights of the partners regarding intellectual property and to act as a mediator in case of disputes
• To foster gender parity in this very male dominated scientific field

While the approach of this project by design mitigates most major concerns regarding security and privacy, the underlying foundations of the platform to be developed need to be secured thoroughly, especially considering the diversity of the local execution platforms on the hospital sites. Important attacker goals include the theft of data on a local level, as well as the theft and manipulation of results, with the inclusion of possible insider attacks. In addition, hospitals also need to be empowered to run experiments on anonymized data in case they are not able to gather the required consent, thus requiring the development of a special anonymization layer. The main objective of WP2 is therefore the definition of a security and privacy architecture based on the cyber risk assessment and legal requirements:

• To develop a risk assessment methodology based on the legal and technical requirements derived from data protection law regulation on the one hand and an in-depth attacker analysis on the other
• To develop suitable Security- and Privacy-Key Performance Indicators (KPIs) for the local execution platforms located inside the hospitals
• To develop an additional layer for adding suitable anonymization to the local execution of the algorithms, in case specific consent could not be gathered
• To select, adapt and further develop mitigation mechanisms, especially considering the detection of malicious algorithms pushed to the local execution platforms, as well as the exfiltration of data besides feature vectors
• To develop a security architecture based on the requirements derived from the risk analysis and the developed mitigation strategies

Software systems developed for molecular diagnostics (MDx) are treated as medical devices and thus have to fulfil specific regulatory requirements, e.g., specific ISO and IEC standards, in order to be approved for clinical use, by e.g. EMA or FDA. At the moment, it is often a better choice to create new compliant software from scratch rather than converting a research software item into MDx software. The main objective of WP3 is therefore guideline development for the software development process, the documentation and the machine learning process:

• To develop a guideline for a standardized software development process for MDx-ready software, which makes a conversion into MDx software feasible
• To develop a documentation guideline for MDx-ready software
• To develop a standardized MDx-ready federal machine learning process which fulfils standards for certification, including data standardization

WP4 will contribute to theoretical and experimental research, design and development of federated interactive learning approaches following “privacy by design and architecture”. Additionally, WP4 will experiment and evaluate explainable AI-approaches in order to make machine learning results transparent, re-traceable, re-enactable and eventually understandable. WP4 is motivated by the fact that unsupervised approaches (WP5) need big data with many training sets; however, in the health domain sometimes we are confronted with a small number of data sets or rare events, where automatic approaches suffer from insufficient training samples. WP4 will thus experiment in the second research stream on making use of a “human-in-the-loop” for solving computationally hard problems, e.g., subspace clustering, protein folding, or k-anonymization of health data, where human expertise can help to reduce an exponential search space through heuristic selection of samples. Consequently, WP4 will contribute to make machine learning results understandable for human experts by applying interactive federated machine learning approaches.

• To create a solid overview on graph parallelism and to form the underlying knowledge base for all our later endeavours
• To foresee the overall topology of a graph that was never seen in its entirety but only implicitly present via its distributed subgraphs and to experiment whether and to what extent such a graph can be thought of being connected in the first place; a sub goal is the exploration of link prediction via node similarity or feature-feature interaction as a necessary pre-processing step
• To shape and compose connection “surfaces” between local subgraphs, which are vital in information propagation behaviour, consequently the goal is to explore whether and to what extent we have to deal with partitions or will be able to take parts (peripheries) of other subgraphs into account as well
• To explore the possibility of mixing strictly separated data with less-sensitive information into “local spheres” enabling privacy-aware federated learning on graphs
• To design, develop and evaluate end-user centred interfaces to enable a) the interaction of humans with the algorithms developed; and b) to enable to re-enact and to re-trace in order to explain and understand the results in the context of the medical problem
• To find the best suitable explanation strategies, i.e. post-hoc and ante-hoc approaches and testing the user interpretation on the demonstrator in order to redesign of the “explanation interface”

The overall objective of the work package is to develop an entire pipeline for the large-scale federated unsupervised learning of biomedical data. The focus is on developing novel methods facilitating de novo endophenotyping of patient cohorts as input for subsequent follow-up analyses, mainly as developed by WP4 and to be integrated in the overall platform developed by WP7.

• To develop an entire federated clustering pipeline (pre-processing, clustering, and cluster evaluation) particularly tailored to the requirements of clustering distributed heterogeneous biomedical datasets
• To develop a novel approach to support the federated clustering of multi-OMICS data with integrated biological network enrichment in order to facilitate systems-medicine based de novo endophenotyping
• To create approaches to intensively evaluate and test the developed approaches and compare their performance to conventional clustering tools on artificial datasets as well as in real-world application scenarios (developed together with WP8) in order to develop usage guidelines and best-practice advices
• To transform the developed and tested prototypes into production ready software with novel data visualization methods and laymen compatible interfaces (Task 4). This also entitles the intensive testing of the software for potential data security issues and the integration in the blockchain user and data-rights management

Most important for the global success of a machine learning platform requiring user consent is the ability of users and data owners to control the data introduced, while allowing data discovery in a privacy-preserving manner. This is especially important in order to integrate as many federated machine learning nodes as possible, while being aware of privacy rights and regulations, especially the General Data Protection Regulation (GDPR) and regulations for medical data. In order to reach these goals, we will conduct research into blockchain-based technologies, especially so-called Byzantine-Fault-Tree (BFT) blockchains, in order to provide user rights management, consent and data discovery mechanisms.

• To provide a local blockchain-mechanism on the individual data holder side in order to track what source data has been used for which analytical instance
• To introduce user-rights management into blockchain mechanisms, where patients can define, which of their sensitive information might be used in federated learning
• To develop mechanisms for selective deletion of sensitive information from the blockchain-mechanism, thus allowing practical consent revocation as demanded by the GDPR
• To provide a global blockchain-based mechanism that allows for the exchange of analysis-relevant meta-information while respecting patient privacy

Key to success of the project is a successful translation into research and clinical practice as well as a wide acceptance by the research community, the clinicians and – most importantly – the patients. To reach this goal we will develop an integrated software platform for a seamless integration of the software solutions developed in WPs 4-6 into an integrated platform framework that is (1) usable for biomedical researchers, clinicians, patients and developers, and (2) extendable by new federated big data technology developed by future users and (bio)medical computer scientists.

• To provide a programming framework (Application Programming Interfaces, short: APIs) allowing to have the methods developed in WPs 4-6 implemented such that they can be integrated into one platform and accessed and controlled in a distributed manner
• To provide an app store to allow the artificial intelligence and data mining community to develop further health project-tailored federated machine learning apps, as well as federated normalization and computational standardization protocols
• To develop, implement and test intuitive user interfaces for patients (data privacy and rights management) and clinicians and hospitals (project management), as well as software developers (registration and testing of new apps)
• To implement automatized measures for evaluating the overall strategy of FeatureCloud by demonstrating that the performance of federated machine learning (in terms of accuracy) is comparable to the performance of traditional cloud-based approaches
• To design and refine the platform such that data security and privacy are optimal (i.e. maximized), and to develop stress test procedures to demonstrate the overall success

The main objective of WP8 is to evaluate the applicability of the FeatureCloud platform in a real-world setting and to successfully validate it on clinical data. To ensure a wide acceptance beyond the research community and to convince clinicians, clinical scientists, and patients to use the FeatureCloud platform, this WP will make sure that it can be used as intuitively as possible while providing all relevant regulatory, scientific and computational features whilst guaranteeing data security in an unprecedented manner.

• To ensure that software programmers are made aware by clinical scientists, biostatisticians, and monitors of all features necessary for usability and to fulfill nationally required reporting features, e.g., serious adverse events, for a phase II/III type clinical study based on MACRO or similar interfaces
• To fine-tune the usability of the FeatureCloud platform during the alpha and beta software developmental phases in real-world clinical settings by providing feedback to the developers of both the platform integration and the user interfaces (Tasks 2 and 3)
• To evaluate the technical performance of the FeatureCloud platform in terms of accuracy and translational power in clinical settings by re-analysing data previously processed in a traditional cloud-based approach and to provide feedback to WPs 4, 5 and 7
• To test and suggest improvements of the user interfaces from different points of view: patients (data privacy and rights management), clinicians and hospitals (trial management, monitoring as well as final and interim data analysis), IT personnel and software developers (registration and testing of new apps)

WP9 will increase the visibility of FeatureCloud by reaching out to the scientific community, industry, patient organisations and other interested or potential stakeholders. A communication plan will be implemented with the following strategic objectives:

• To make FeatureCloud known to the scientific community and the public
• To disseminate the results to the scientific community in the academic and healthcare sectors and foster interaction and exchange with the scientific community and the public
• To train and prepare the next generation of software developers, clinicians, patients and the next generation of medicine students that grow into a world of artificial-intelligence-supported precision medicine
• To identify and valorise the intellectual property rights (IPR) generated within WP1-8 and initiate next steps for the foundation of a spin-out company to ensure future support and developments of the FeatureCloud platform and privacy-preserving machine learning in general

The objective of this WP is to ensure that all procedures and partners comply with the ethics requirements developed and set out by this work package.