Federated Representation Adaptation for Harmonizing Cross-Border Healthcare Records Under Privacy Constraints

The high demand for healthcare analytics that maintains patient confidentiality will necessitate a solution that satisfies the requirements of the GDPR and HIPAA. FL substantially lessens the risks associated with centralized data aggregation and facilitates collaborative model training across multiple institutions while protecting local data. Attacks that utilize local updates alongside gradient inversion to unveil sensitive data from shared gradients are directed at traditional FL ecosystems. To combat these challenges, this paper proposes a new framework, FedSplit-Adpt framework that develops cross-border healthcare data harmonization within privacy constraints using split learning with vertical data partitioning. The framework answers the issue of merging heterogeneous health data from multiple institutions spread across various nations so that the data features are vertically partitioned, with each institution having a different, smaller subset of patient information. Split learning is leveraged to partition the neural network between client-side and server-side, preserving the privacy of patient information at the point of use, while transmitting intermediate activations for further computing, thus assuring compliance with privacy regulations. Vertical data partitioning allows the model to learn from entire patient profiles and amalgamate characteristics from many institutions, knowing that sensitive data is still protected. Institutions only share encrypted intermediate outputs, since encryption is done using homomorphic encryption for added security. During collaborative training, the feasible optimized intermediate representations are implemented to build a global model that can unify data from many healthcare settings, using disparate terminologies, feature distributions, and medical standards. The results demonstrate that the FedSplit-Adapt framework standardizes healthcare data from external settings with 96.4% success and 20% data interoperability, comp...