Jie Xu1, Yu Zhang2, Huamin Yu3, Bo Lin4,5,6, Dejian Wang6, Hong Yuan3, Bin Hu7, Jun Jiang2, Peng Xiang8, Te Lin1, Huizhe Lu1, Guiying Zhang6. 1. Information Center, the Second Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou, China. 2. Department of Cardiology, the Second Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou, China. 3. Department of Cardiovascular Medicine, the First People's Hospital of Linping, Hangzhou, China. 4. College of Computer Science and Technology, Zhejiang University, Hangzhou, China. 5. Innovation Centre for Information, Binjiang Institute of Zhejiang University, Hangzhou, China. 6. Department of R&D, Hangzhou Healink Technology, Hangzhou, China. 7. Department of Cardiovascular Medicine, Jiande First People's Hospital, Hangzhou, China. 8. Department of Vasculocardiology, Institute of Computing Innovation, Zhejiang University, Hangzhou, China.
Abstract
Background: Multicenter medical research is becoming a new trend. However, interagency data privacy security protection has become a major bottleneck of multicenter research. Therefore, to overcome this privacy protection issue, the aim of the present study was to apply a self-developed privacy-preserving machine learning framework for researchers who can build models on medical data from multiple sources, while providing privacy protection for both sensitive data and the learned model. Methods: Based on Arya, a novel privacy computing platform developed by Healink, we constructed a privacy-preserving federated learning (FL) model using the fully connected neural network with datasets from 2-3 individual medical institutions. In the dataset, 80% of records were used for joint modeling on acute myocardial infarction (AMI) diagnosis. Modeling efficacy was evaluated with the remaining 20% of records. As the control, 1,500 medical records from 1 medical institution were used for single-center modeling and efficacy evaluation. During the process, the original data were still kept in individual hospital without moving or transferring out of the hospitial. The diagnostic efficacy (sensitivity, positive predictive value, and accuracy) was evaluated. Results: Our privacy-preserving FL model gives reliable AMI diagnostic efficacy. Three-center modeling (79% sensitivity, 88% positive predictive value, and 82.3% accuracy) and two-center modeling (77.8% or 77.6% sensitivity, 86.7% or 85.30% positive predictive value, and 81% or 79.7% accuracy) achieved relative high diagnostic efficacy; and single-center modeling achieved relative low diagnostic efficacy (76% sensitivity, 84.7% positive predictive value, and 79% accuracy). Conclusions: The Arya privacy computing platform is efficient and practical for the FL model, which could promote multicenter medical research securely without sacrificing diagnostic efficacy. 2022 Annals of Translational Medicine. All rights reserved.
Background: Multicenter medical research is becoming a new trend. However, interagency data privacy security protection has become a major bottleneck of multicenter research. Therefore, to overcome this privacy protection issue, the aim of the present study was to apply a self-developed privacy-preserving machine learning framework for researchers who can build models on medical data from multiple sources, while providing privacy protection for both sensitive data and the learned model. Methods: Based on Arya, a novel privacy computing platform developed by Healink, we constructed a privacy-preserving federated learning (FL) model using the fully connected neural network with datasets from 2-3 individual medical institutions. In the dataset, 80% of records were used for joint modeling on acute myocardial infarction (AMI) diagnosis. Modeling efficacy was evaluated with the remaining 20% of records. As the control, 1,500 medical records from 1 medical institution were used for single-center modeling and efficacy evaluation. During the process, the original data were still kept in individual hospital without moving or transferring out of the hospitial. The diagnostic efficacy (sensitivity, positive predictive value, and accuracy) was evaluated. Results: Our privacy-preserving FL model gives reliable AMI diagnostic efficacy. Three-center modeling (79% sensitivity, 88% positive predictive value, and 82.3% accuracy) and two-center modeling (77.8% or 77.6% sensitivity, 86.7% or 85.30% positive predictive value, and 81% or 79.7% accuracy) achieved relative high diagnostic efficacy; and single-center modeling achieved relative low diagnostic efficacy (76% sensitivity, 84.7% positive predictive value, and 79% accuracy). Conclusions: The Arya privacy computing platform is efficient and practical for the FL model, which could promote multicenter medical research securely without sacrificing diagnostic efficacy. 2022 Annals of Translational Medicine. All rights reserved.
Authors: Jared L Katzman; Uri Shaham; Alexander Cloninger; Jonathan Bates; Tingting Jiang; Yuval Kluger Journal: BMC Med Res Methodol Date: 2018-02-26 Impact factor: 4.615