Federated Learning-based Healthcare Analytics for Privacy-Preserving Patient Data Sharing

Le résumé fourni par la source

Abstract The new federated learning healthcare analytics framework is necessary because it helps solve the major challenges posed by traditional centralized machine learning approaches to the healthcare system. The traditional way used in machine learning is to gather sensitive patient information into one place to perform analysis. Because there is a central repository of information in a healthcare environment, there is an increased risk of a data breach occurring, an unauthorized access to sensitive patient data, and insufficient compliance with strict federal regulations, including HIPAA, and GDPR. Therefore, the limitations of the centralized systems make these traditional methods incompatible with today's healthcare environments because data privacy and data security are the highest priorities.The introduction of a federated learning validation system allows for a strong privacy-preserving distributed model training system in which many medical facilities can work together to develop a global healthcare model without having to share raw patient data. Participants will work cooperatively to create a global healthcare model by only sharing model updates rather than their original data with any other participant. The federated learning distributed model training system reduces the amount of exposure of sensitive patient data to other participants and is mitigated by not violating the privacy rights of any participant. Additionally, since each participant entity has the ability to retain ownership and maintain confidentiality of each virtual machine database, the framework will ensure regulatory compliance with their respective regulatory authorities. Incorporation of complex security mechanisms adds another layer of security to this framework's reliability. For example, secure aggregation allows the individual models updates to be unaccessed or reconstructed by a third party, whereas differential privacy uses added noise to prevent individuals from being identified from their EHR data. Furthermore, lightweight encryption adds another level of data security during transmission, thus, increasing the likelihood of preventing cybersecurity incidents and/or guess attacks. Through the experimental evaluation of this proposed framework utilizing simulated electronic health records (EHR) and medical imaging datasets, it has been shown that federated models exhibit performance characteristics comparable to centralized models as determined by prediction accuracy. Moreover, the use of this framework significantly reduces any privacy threat through the reduction of class membership inference and re-identification attacks. The balance of privacy preservation and prediction accuracy provides further evidence of the practicality of this approach for providing improved data security while enabling greater collaborative practices among healthcare organizations. Lastly, the utilization of this framework may further increase the number of organizations that can participate in joint/pooled efforts towards the creation of an intelligence model resulting in enhanced generalizability and enhanced healthcare analytics. This feature is especially beneficial in situations where heterogeneity of data is required to make accurate diagnoses and predictions (e.g., disease detection or medical imaging analyses).

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.