Source-linked AI summary
Privacy-Preserving Technology to Help Millions of People: Federated Prediction Model for Stroke Prevention
Ce Ju, Ruihui Zhao, Jichao Sun, Xiguang Wei, Bo Zhao, Yang Liu, Hongshan Li, Tianjian Chen, Xinwei Zhang, Dashan Gao, Ben Tan, Han Yu, Chuning He, Yuan Jin
TL;DR
Privacy concerns around distributed hospital EHR data limit aggregation for stroke-prediction training. The paper uses federated averaging to train across hospitals without moving patient data, and reports that federated AUC means were almost the same as centralized training, with slightly larger variance.
Problem
Privacy concerns around distributed EHR data limit aggregating patient records for centralized stroke-prediction training.
Source: §1 · p. 1Method
The system uses federated averaging to aggregate hospital model updates while patient data remain in each client’s database.
Source: §3.2 · pp. 2–3, §1 · p. 2Results
The federated model’s AUC mean was almost the same as centralized training, with a little bigger variance.
Source: §4 · p. 3Takeaways & Limitations
The authors describe the work as a privacy-preserving stroke-prediction pipeline with social value.
Source: §5 · p. 3
Abstract
from arXiv · showhide
Prevention of stroke with its associated risk factors has been one of the public health priorities worldwide. Emerging artificial intelligence technology is being increasingly adopted to predict stroke. Because of privacy concerns, patient data are stored in distributed electronic health record (EHR) databases, voluminous clinical datasets, which prevent patient data from being aggregated and restrains AI technology to boost the accuracy of stroke prediction with centralized training data. In this work, our scientists and engineers propose a privacy-preserving scheme to predict the risk of stroke and deploy our federated prediction model on cloud servers. Our system of federated prediction model asynchronously supports any number of client connections and arbitrary local gradient iterations in each communication round. It adopts federated averaging during the model training process, without patient data being taken out of the hospitals during the whole process of model training and forecasting. With the privacy-preserving mechanism, our federated prediction model trains over all the healthcare data from hospitals in a certain city without actual data sharing among them. Therefore, it is not only secure but also more accurate than any single prediction model that trains over the data only from one single hospital. Especially for small hospitals with few confirmed stroke cases, our federated model boosts model performance by 10%~20% in several machine learning metrics. To help stroke experts comprehend the advantage of our prediction system more intuitively, we developed a mobile app that collects the key information of patients' statistics and demonstrates performance comparisons between the federated prediction model and the single prediction model during the federated training process.
1 Introduction
The section presents privacy concerns around health-care data as a barrier to using aggregated data for stroke prediction. It proposes federated learning as a way to train collaboratively without sharing hospital data.
1 Introduction
- Federated learning can jointly train a machine-learning model across clients without actually sharing their data.The passage says the federated model performs more accurately and securely than the traditional statistical model in medical research.Source: §1 · pp. 1–2
- The section connects health-care data privacy risks with the need for privacy-preserving AI mechanisms.It notes that health-care data include clinical metrics and health and wellness information, and cites concerns that big-data use may lead to privacy violations.Source: §1 · p. 1
- The authors develop a privacy-preserving technology for stroke prediction to address privacy-related scalability concerns.The federated version builds on Tencent’s prediction model and WeBank’s privacy-preserving framework.Source: §1 · p. 2
2 Related Work
The related work defines federated learning as collaborative machine learning in which client data remain in their databases, and reviews examples across several health-care domains.
2 Related Work
- Federated learning lets clients collaborate on a machine-learning problem while their data remain in their databases [Kairouz et al., 2019].The cited health-care work includes decentralized EHR learning, multi-institutional medical imaging, differential privacy for brain-tumor segmentation, and brain-computer interface models.Source: §2 · p. 2
- Prior work applied federated approaches to decentralized EHR analysis and multi-institutional medical imaging [Sheller et al., 2018].Brisimi et al. developed an iterative cluster Primal-Dual Splitting algorithm; Sheller et al. introduced federated learning for medical-image collaboration.Source: §2 · p. 2
- Other studies examined differential privacy for brain-tumor segmentation and distributed brain-computer interface training across heterogeneous datasets [Li et al., 2019; Gao et al., 2019; Ju et al., 2020].Source: §2 · p. 2
3 Methodology
The methodology prepares hospital clinical data and features, trains stroke-risk classifiers through federated averaging, and provides a system for monitoring the federated process.
3.1 Workflow of the Stroke Prediction Model
- Figure 1 is identified as the workflow for the stroke prediction model.Source: §3.1 · p. 2
Data Preprocessing
- The system cleans and normalizes hospital patient data into structured clinical data, then extracts model features into a disease feature library stored in private clouds.The preprocessing uses medical knowledge and AI techniques such as natural language processing and relation networks.Source: Data Preprocessing · p. 2
Missing Data Imputation
- Missing clinical-data entries are imputed with column means or medians, then rounded to the nearest discrete value.The passage motivates imputation as a response to omissions that can make predictions inaccurate.Source: Missing Data Imputation · p. 2
Feature Selection
- Medical experts manually select 119 stroke-relevant model features based on risk factors analyzed in medical studies [Chambless et al., 2004; Lumley et al., 2002; Vokó et al., 2004].Examples include systolic blood pressure, antihypertensive therapy, and prior coronary disease.Source: Feature Selection · p. 2
Prediction Classifier
- The classifier outputs stroke-risk scores using logistic regression, random forest, XGBoost, or neural networks.The classifier is established on model features from the feature library.Source: Prediction Classifier · p. 2
3.2 Federated Prediction Model
- The server aggregates hospital model updates, returns the updated global model, and supports asynchronous training with any number of clients and local gradient iterations per round.Patient data remain in each hospital’s private cloud during the process.Source: §3.2 · pp. 2–3
3.3 Monitoring and Visualization System
- The monitoring and visualization system tracks data flow among clients and the server and displays patient statistics, histograms, and federated-versus-single-model AUC performance.The interface is a mini-program for hospital doctors.Source: §3.3 · p. 3
4 Experimental Analysis
Federated training improves stroke prediction over single-hospital models, particularly for small hospitals with few confirmed cases. In the citywide evaluation, its mean AUC was nearly the same as centralized training, with slightly greater variance.
4 Experimental Analysis
- 10% ∼20% in AUC: federated training improves on single-hospital models, especially at small hospitals with few confirmed cases.The evaluation used hospitals in a city in Central China.Source: §4 · p. 3
- Federated training’s mean AUC was almost the same as centralized training, with slightly greater variance.Hospital A accounted for around 50% of patients, so its performance dominated joint training and was close to centralized training.Source: §4 · p. 3
- The evaluation used a 3-layer neural-network classifier, five experiments, and a fixed test set sampled from patients across the city.Hospitals A, B, and C had 3%–4% clinically confirmed stroke patients; hospitals D and E had less than 1%.Source: §4 · p. 3
5 Discussion
The authors describe their work as a privacy-preserving mechanism intended to boost stroke-prediction performance. They characterize the resulting healthcare pipeline as the first of its kind worldwide and emphasize its social value.
5 Discussion
- The authors characterize their privacy-preserving stroke-prediction pipeline as the first such healthcare pipeline worldwide and say it has profound social value.They state that the mechanism is designed to boost prediction-model performance.Source: §5 · p. 3