Secure distributed multiple imputation enables missing data inference for private data proprietors

TL;DR

This paper introduces a secure method for imputing missing data in private health records across multiple institutions without compromising privacy.

Contribution

The novel contribution is a provably secure distributed imputation framework using secure multiparty computation for collaborative analysis of private EHRs.

Findings

The proposed method achieves practical runtimes and accuracy comparable to non-secure imputation techniques.

It enables collaborative studies on incomplete, private datasets without centralized data pooling.

The framework improves classification of high-risk ICU patient outcomes using real-world data.

Abstract

Scattered between many healthcare providers across the US, Electronic Health Records (EHR) are extensively used for research purposes. Collaboration and sharing of EHRs between multiple institutions often provide access to more diverse datasets and a chance to conduct comprehensive studies. However, these collaboration efforts are usually hindered by privacy issues that render the pooling of such data at a centralized database impossible. Furthermore, EHRs are often incomplete and require statistical imputation prior to the study. To enable collaborative studies on top of incomplete, private EHRs, here we provide a provably secure solution built with secure multiparty computation (SMC) that provides practical runtimes and accuracy on par with the state-of-the-art, non-secure equivalents. Our solution enables the utilization of distributed datasets as a whole to impute the missing data and conduct collective studies between non-trusting private data proprietors. We demonstrate its effectiveness on various synthetic and real-world datasets, and show that our solution can significantly improve the classification of high-risk patient outcomes during ICU admission.

Figures