Federated Learning of Predictive Models from Real Data on Diabetic Patients

Abstract

Today Federated Learning is gaining more and more notoriety in the medical field, especially since it can guarantee the privacy of sensitive data, unlike Machine Learning. Furthermore, training a model on a single patient does not always guarantee acceptable results as we are different from each other. Starting from this assumption, the idea would therefore be to create a system that allows diabetic patients to share their data on the cloud (also using the Cloud/Edge Continuum technology and clearly going to respect all the constraints related to privacy by differentiating the use of the public cloud from the private one), collected via smartwatch, and, through Federated Learning, to train a centralized model that allows them to go to predict when one of the patients may need to inject insulin into his/her body. In particular, this document represents a part of a larger work in which various Time Series Analysis and Deep Learning algorithms were tested on real data from a diabetic patient, making predictions to understand, among those analyzed, which was the best algorithm on a single patient and then generalize by applying a multi-client approach.