Weight aggregation methods for federated learning in healthcare - A comparative empirical analysis

Ensuring the safety and confidentiality of patient data is critical to the healthcare industry. The transmission of raw data between satellite and central servers within healthcare conglomerates represents a weak link in the chain of protected health information (PHI). Federated Learning approaches that do not transmit raw data minimize this risk. In this paper, we present an approach to improve the prediction accuracy of patient imaging at individual hospitals in a Federated Learning (FL) setting without sharing raw data. The learning focuses on sharing classifier weights from local distributed learning with private data. We also present a comparative empirical analysis of several approaches to aggregate the weight parameters obtained from distributed learning. The approach assumes that each client has access to private datasets while a shared common dataset at the server level is accessible to all clients. The model is trained at the server level using the shared dataset which is further personalized for each client by training it on their private data. We investigated six approaches— Federated Averaging, Federated Client Averaging, Federated Averaging Scaled Minimum Shifted, Federated Client Averaging Minimum Shifted, Federated Averaging Scaled Maximum Shifted, and Federated Client Averaging Maximum Shifted—to aggregate the updated client weights at the server level. Our experiments used the Modified National Institute of Standards and Technology (MNIST) dataset to evaluate the performance of the resulting model at the server level. The Federated Client Averaging Minimum Shifted approach showed better performance in our experiments. Our results indicate that Federated Client Averaging Minimum Shifted is a promising option for FL in distributed and privacy-sensitive environments.