Federated Machine Learning, Privacy-Enhancing Technologies, and Data Protection Laws in Medical Research: Scoping Review.

BACKGROUND: The collection, storage, and analysis of large data sets are relevant in many sectors. Especially in the medical field, the processing of patient data promises great progress in personalized health care. However, it is strictly regulated, such as by the General Data Protection Regulation (GDPR). These regulations mandate strict data security and data protection and, thus, create major challenges for collecting and using large data sets. Technologies such as federated learning (FL), especially paired with differential privacy (DP) and secure multiparty computation (SMPC), aim to solve these challenges. OBJECTIVE: This scoping review aimed to summarize the current discussion on the legal questions and concerns related to FL systems in medical research. We were particularly interested in whether and to what extent FL applications and training processes are compliant with the GDPR data protection law and whether the use of the aforementioned privacy-enhancing technologies (DP and SMPC) affects this legal compliance. We placed special emphasis on the consequences for medical research and development. METHODS: We performed a scoping review according to the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews). We reviewed articles on Beck-Online, SSRN, ScienceDirect, arXiv, and Google Scholar published in German or English between 2016 and 2022. We examined 4 questions: whether local and global models are "personal data" as per the GDPR; what the "roles" as defined by the GDPR of various parties in FL are; who controls the data at various stages of the training process; and how, if at all, the use of privacy-enhancing technologies affects these findings. RESULTS: We identified and summarized the findings of 56 relevant publications on FL. Local and likely also global models constitute personal data according to the GDPR. FL strengthens data protection but is still vulnerable to a number of attacks and the possibility of data leakage. These concerns can be successfully addressed through the privacy-enhancing technologies SMPC and DP. CONCLUSIONS: Combining FL with SMPC and DP is necessary to fulfill the legal data protection requirements (GDPR) in medical research dealing with personal data. Even though some technical and legal challenges remain, for example, the possibility of successful attacks on the system, combining FL with SMPC and DP creates enough security to satisfy the legal requirements of the GDPR. This combination thereby provides an attractive technical solution for health institutions willing to collaborate without exposing their data to risk. From a legal perspective, the combination provides enough built-in security measures to satisfy data protection requirements, and from a technical perspective, the combination provides secure systems with comparable performance with centralized machine learning applications.

Privacy-aware multi-institutional time-to-event studies.

Clinical time-to-event studies are dependent on large sample sizes, often not available at a single institution. However, this is countered by the fact that, particularly in the medical field, individual institutions are often legally unable to share their data, as medical data is subject to strong privacy protection due to its particular sensitivity. But the collection, and especially aggregation into centralized datasets, is also fraught with substantial legal risks and often outright unlawful. Existing solutions using federated learning have already demonstrated considerable potential as an alternative for central data collection. Unfortunately, current approaches are incomplete or not easily applicable in clinical studies owing to the complexity of federated infrastructures. This work presents privacy-aware and federated implementations of the most used time-to-event algorithms (survival curve, cumulative hazard rate, log-rank test, and Cox proportional hazards model) in clinical trials, based on a hybrid approach of federated learning, additive secret sharing, and differential privacy. On several benchmark datasets, we show that all algorithms produce highly similar, or in some cases, even identical results compared to traditional centralized time-to-event algorithms. Furthermore, we were able to reproduce the results of a previous clinical time-to-event study in various federated scenarios. All algorithms are accessible through the intuitive web-app Partea (https://partea.zbh.uni-hamburg.de), offering a graphical user interface for clinicians and non-computational researchers without programming knowledge. Partea removes the high infrastructural hurdles derived from existing federated learning approaches and removes the complexity of execution. Therefore, it is an easy-to-use alternative to central data collection, reducing bureaucratic efforts but also the legal risks associated with the processing of personal data to a minimum.