Tang Zhaohui, Van Nguyen Thi Phuoc, Yang Wencheng, Xia Xiaoyu, Chen Huaming, Mullens Amy B, Dean Judith A, Osborne Sonya R, Li Yan
School of Mathematics, Physics and Computing, Centre for Health Research, University of Southern Queensland, Toowoomba Campus, QLD, Australia.
Department of Information Technology, Thanh Do University, Hanoi, Vietnam.
Digit Health. 2024 Nov 21;10:20552076241298425. doi: 10.1177/20552076241298425. eCollection 2024 Jan-Dec.
Applying and leveraging artificial intelligence within the healthcare domain has emerged as a fundamental pursuit to advance health. Data-driven models rooted in deep learning have become powerful tools for use in healthcare informatics. Nevertheless, healthcare data are highly sensitive and must be safeguarded, particularly information related to sexually transmissible infections (STIs) and human immunodeficiency virus (HIV).
We employed federated learning (FL) in combination with homomorphic encryption (HE) for STI/HIV prediction to train deep learning models on decentralized data while upholding rigorous privacy. The dataset included 168,459 data entries collected from eight countries between 2013 and 2018. The data for each country was split into two groups, with 70% allocated for training and 30% for testing. Our strategy was based on two-step aggregation to enhance model performance and leverage the area under the curve (AUC) and accuracy metrics and involved a secondary aggregation at the local level before utilizing the global model for each client. We introduced a dropout approach as an effective client-side solution to mitigate computational costs.
Model performance was progressively enhanced from an AUC of 0.78 and an accuracy of 74.4% using the local model to an AUC of 0.94 and an accuracy of 90.7% using the more advanced model.
Our proposed model for STI/HIV risk prediction surpasses those achieved by local models and those constructed from centralized data sources, highlighting the potential of our approach to improve healthcare outcomes while safeguarding sensitive patient information.
在医疗保健领域应用和利用人工智能已成为促进健康的一项基本追求。基于深度学习的数据驱动模型已成为医疗保健信息学中强大的工具。然而，医疗保健数据高度敏感，必须加以保护，特别是与性传播感染（STIs）和人类免疫缺陷病毒（HIV）相关的信息。
我们将联邦学习（FL）与同态加密（HE）相结合用于性传播感染/艾滋病毒预测，以便在分散的数据上训练深度学习模型，同时严格维护隐私。该数据集包括2013年至2018年期间从八个国家收集的168459条数据记录。每个国家的数据分为两组，70%用于训练，30%用于测试。我们的策略基于两步聚合，以提高模型性能并利用曲线下面积（AUC）和准确率指标，并且在为每个客户端使用全局模型之前在本地级别进行二次聚合。我们引入了一种随机失活方法作为有效的客户端解决方案来降低计算成本。
模型性能从使用本地模型时的AUC为0.78和准确率为74.4%逐步提高到使用更先进模型时的AUC为0.94和准确率为90.7%。
我们提出的性传播感染/艾滋病毒风险预测模型超过了本地模型以及从集中数据源构建的模型，突出了我们的方法在保护敏感患者信息的同时改善医疗保健结果的潜力。
