An Investigation of Federated Learning Strategies for Disease Diagnosis
Abhay Das, Aishwarya Krishnadas, Vaishakh S Krishnan, Avani Farida, Greeshma Sarath · 2024
Traditional healthcare systems utilize centralized approaches for building machine-learning models for disease diagnosis. It requires sharing raw data to the centralized server, which is practically difficult due to patient’s privacy. This leads to significant vulnerabilities and challenges. As a result, switching to approaches for scalable and distributed AI at the network edge for privacy-preserving smart healthcare applications is essential. Deep learning models need to train on data, on any nodes, in a privacy-preserving manner, but not give up the distinction of patients. This paper aims to investigate the application of FL in providing health diagnosis while enhancing the patient’s data privacy, by adopting the Flower platform. It employs the following four algorithms for federated learning that can facilitate collaborative model training with shared patient data: FedAvg, FedProx, FedAdagrad, and FedAdam. To evaluate FL’s effectiveness, deep learning models trained on decentralized COVID-19 and Pneumonia datasets are used from different sources. FL methods were compared to a Centralized Convolutional Neural Network (CNN) to check the performance. The analysis shared in this paper shows that implementing FL contributes to establishing the subject’s accurate diagnostics concerning the patient’s data protection in the healthcare industry.