Privacy-Protecting Analytic and Data-Sharing Methods for Large Electronic Health Data Networks

• Activity Number: 125
• Type: Topic Contributed
• Date/Time: Monday, August 10, 2015 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Darren Toh*
• Companies: Harvard Medical School/Harvard Pilgrim Health Care Institute
• Keywords: Distributed methods ; Comparative effectiveness research ; Pharmacoepidemiology ; Propensity scores ; Active surveillance ; Sentinel

Abstract

Electronic healthcare databases are widely used in comparative effectiveness and safety research of medical treatments. For privacy and practical reasons, it is often necessary to minimize sharing of individual-level data in multi-site studies that analyze these databases. This creates analytic challenges as individual-level information is often needed to adjust for biases inherent in observational studies. In this session, we describe two studies conducted within large electronic healthcare data networks that employed a suite of privacy-protecting methods to perform complex statistical analysis with only summary-level information. These methods exploit the properties of confounder summary scores (e.g., propensity scores, disease risk scores), which condense information from a large number of covariates into non-identifiable measures. We discuss the strengths and limitations of these methods, and the feasibility and challenges of applying them in real-world distributed data networks. We also discuss future directions in this research area, including the use of other privacy-protecting methods such as distributed regression in these distributed data networks.