International conference on Health Informatics (HEIN 2023)
March 11-12, 2023, Virtual Conference
Personalized Progressive Federated Learning with Leveraging Client-specific Vertical Features: Model Development and Validation
Tae Hyun Kim, Won Seok Jan, Sun Cheol Heo, Min Dong Sung and Yu Rang Park, Department of Biomedical Systems Informatics, Yonsei University College of Medicine, Seoul, South Korea
Federated learning (FL) has been used for model building across distributed clients. However, FL cannot leverage vertically partitioned features to increase the model complexity. In this study, we proposed a personalized progressive federated learning (PPFL) model, which is a multi-model PFL approach that allows the leveraging of vertically partitioned client-specific features. The performance of PPFL was evaluated using two datasets: the Physionet Challenges 2012 dataset and a real-world dataset composed of eICU data and the Severance Hospital, Seoul, South Korea. We compared the performance of inhospital mortality and length of stay prediction between our model and the FedAvg, FedProx, and local models. The PPFL showed an accuracy of 0.849 and AUROC of 0.790 in in hospital mor-tality prediction, which are the highest scores compared to client-specific algorithm. For length-of-stay prediction, PPFL also showed an AUROC of 0.808 in average which was the highest among all comparators.