Privacy Preserving Federated Learning for Decentralized Electronic Health Records in Smart Cities

Abstract

The emerging trends of smart healthcare systems are creating immense amounts of electronic health records (EHRs) and providing novel opportunities in medical analytics based on data. Nonetheless, centralized machine learning methods need to consolidate sensitive patient information, and this poses serious privacy and security issues. To overcome these issues, this paper suggests a privacy-conserving federated learning model of decentralized electronic health record systems in smart city healthcare systems. The suggested model allows a joint model training between various healthcare organizations without transferring raw patient data. The strategy combines federated learning with the mechanism of differential privacy and secure aggregation to make sure that sensitive medical information is not disclosed in the process of distributed model training. The experimental analysis proves that the suggested framework has a competitive predictive performance and strong privacy protection. The centralized learning performance was found to be comparable to the model, with an accuracy of 91.0, precision of 90.3, recall of 89.4, and an F1-score of 89.8, making the model much better for data confidentiality. The convergence analysis shows that model training is stable with 80-100 communication rounds, and communication overhead is not too high as the number of participating hospitals increases up to 20 institutions. A further case in support of the integration of privacy-preserving mechanisms is an ablation study, which confirms only a slight decrease in the predictive performance but increases the data security. The suggested framework is a successful remedy to secure collaborative healthcare analytics in decentralized smart city settings.