[4Xin1] Poster session 2

[4Xin1-09] Comparison of Communication Cost and Accuracy in Federated Learning and Data Collaboration Analysis

Assuming Integrated Data Analysis of University of Tsukuba Hospital and Tsukuba City Hall

〇Akihiro Toyoda¹, Go Uchitachimoto², Masayuki Kojima², Rina Kagawa³, Takashi Oyama⁴, Akira Imakura^5,6, Tetsuya Sakurai^5,6, Yukihiko Okada^5,6 (1.Univ. of Tsukuba, The College of Policy and Planning Science, 2. Univ. of Tsukuba, Master's Program in Service Engineering, 3.Univ. of Tsukuba, Institute of Medicine, 4.Tsukuba City, Health Department, National Health Insurance Division, 5.Univ. of Tsukuba, Institute of Systems and Information Engineering, 6.Univ. of Tsukuba, Center for Artificial Intelligence Research)

Keywords:Data collaboration analysis, Federated Learning, Distributed data, Privacy preserving, Medical data

Federated Learning (FL) has been studied assuming a constant Internet connection, but often sensitive distributed data cannot always be connected. Even in this situation, FL can perform integrated data analysis by using different communication methods. Data Collaboration (DC) analysis has been proposed for integrated data analysis without a permanent Internet connection or sharing of private data. This study constructs a machine learning model to predict fasting blood glucose using medical data from the University of Tsukuba Hospital and Tsukuba City Hall, where raw data cannot be shared, and a constant Internet connection is not available. Accuracy and computational cost of following three methods are compared: individual analysis at each institution, FL without a constant Internet connection, and DC. Results show that FL and DC generally improve prediction accuracy compared to individual analysis. DC is twice as computationally intensive as FL but achieves the same accuracy in 1/3 of the time.