Asynchronous Blockchain-based Privacy-preserving Training Framework for Disease Diagnosis

With the emerging artificial intelligence technology, especially machine learning and deep learning, an increasing number of healthcare institutions have designed and implemented various data-driven analysis tools and models to assist disease diagnosis. However, it is difficult for a single healthcare institution to collect sufficient medical records to support a sophisticated automatic disease recognition model, especially for some rare diseases, which urges the collaboration among different medical institutions and hospitals. The current mainstream solutions, like centralized distributed machine learning, heavily rely on a central server, which may be vulnerable to attackers or even malicious itself. Meanwhile, data privacy concerns make hospitals reluctant to share their patients' records with others. To solve these issues, in this work, we utilize the blockchain to build a decentralized privacy-preserving cross-institution disease classification framework, called Health-Chain. Specifically, we combined differential privacy and pseudo-identity mechanism to protect data privacy in distributed stochastic gradient descent (SGD) algorithm. Meanwhile, we equip gradient delay compensation to address the asynchronous issues in the decentralized blockchain-based learning system. In the experiments, we implement our Health-Chain in two popular disease recognition tasks, breast cancer diagnosis, and ECG arrhythmia classification, and demonstrate the efficiency and effectiveness of the proposed framework.