Study On Multicenter Clinical Data Analysis Based On Homomorphic Encryption

In order to carry out high-quality medical research,it is often necessary to collect clinical data from multiple medical institutions to obtain more repeatable,general and novel research results.However,since the information of patients is highly sensitive,data sharing in multicenter medical research has the risk of privacy leakage,which may lead to serious losses to the exposed patients and their data providers.Therefore,it is of great significance to carry out multicenter clinical data analysis on the premise of avoiding sensitive information leakage of any patients.Most existing privacypreserving multicenter clinical data analysis methods cannot take into account high security and low accuracy loss at the same time,and the calculation and transmission consumption are very large,making them unable to meet the needs of real-world clinical applications.Therefore,this paper proposed a high-performance,high-precsion,high-security multicenter clinical data analysis framework,which opened up a new mode for the secure use of clinical data.The main innovation points are as follows:A multicenter clinical data analysis framework based on homomorphic encryption was proposed.Since traditional homomorphic encryption technique cannot be applied to multicenter clinical data analysis,a threshold homomorphic encryption scheme that allowed floating-point number calculation was proposed.Meanwhile,the key generation process was improved,making the scalability and accuracy of the framework better.The proposed framework can resist conspiracy attacks,and provide a strong security guarantee for research data and analysis results.In terms of security,it is significantly better than the existing privacy-preserving multicenter medical research frameworks,which was conducive to the sharing of clinical data across different medical institutions.A logistic regression analysis scheme based on threshold homomorphic encryption was proposed and implemented,which achieved the privacy protection from model training to model evaluation,and provided medical researchers with a truly practical privacy-preserving multicenter classification analysis tool.As for model training,a parallel training strategy based on batch encoding was proposed,making its training efficiency much better than existing similar schemes.A Cox regression analysis scheme based on threshold homomorphic encryption was proposed and implemented,which was applicable to both horizontally and vertically partitioned dataset.It had excellent security,scalability,accuracy of results,and could simultaneously predict survival time and evaluate the importance of the variable.The proposed scheme is significantly better than the latest similar research in terms of computational efficiency,which is more suitable to the clinical data analysis with large dataset.This study eliminated the privacy concerns of medical institutions,solved the practicability issue of homomorphic encryption in real-world clinical applications,was of great significance to real-world drug discovery and other aspects.