Predicting the Prognosis of Stroke Patients Based on Personalized Federated Learning

Abstract

In recent years, the incidence of stroke has been increasing and showing a trend of younger people. Based on the distributed stroke risk assessment modeling scenario, this paper solves the problems of insufficient data and difficult data sharing in medical institutions through federated learning. Considering the features of structured data, the proposed algorithm takes the non-neural network model as the base model, and combines bagging and gradient boosting algorithms to achieve model updating and aggregation. This paper also proposes the model pruning method to realize the personalization of each participant’s model and reduces the data transmission cost of the algorithms by separating the weight matrix of the model and the model parameters. Experiments show that the proposed method greatly outperforms existing baseline approaches according to the predictive results, and the accuracy of the personalized model and the global model in the International Stroke Trial (IST) dataset reaches 78.20% and 76.85%, respectively, which has broader application scenarios.