Advanced theory and methods for the de-identification of small cohorts, complex and composed health data

点击翻译按钮获取中文摘要

The demand for health data for research and public health purposes has never been so great, from the growth in clinical trials transparency initiatives, Electronic Medical Records being used to build learning healthcare systems, the development of real world evidence databases that integrate health data from multiple sources and used for observational research, and a desire to link to wearables and other monitoring devices. The collection of data is coming from providers, payers, employers, wellness programs, and even patients themselves, with increasing collaboration between academic institutions, provider organizations, health care systems and life sciences companies.***We have made many critical achievements in our privacy research which represent improvements over the existing body of work. As health data uses evolve and the nature of health data that is being shared also evolves, there are important areas that require more study:*** - One of the best ways to allow the sharing of data for secondary purposes is to de-identify it. A key part of de-identification is the estimation of re-identification risk. Small cohorts pose a particular challenge to estimate re-identification risk. We will develop suitable estimators for small data sets. This will be important in developing successful models for clinical trials data sharing and studies on rare diseases and conditions.***- With the growing number of sources of data, there is more demand to join data sets for building real world evidence databases. Often the individual de-identified data are being linked without consideration of the potential increase in re-identification risk. There is a need to develop a composition theory around re-identification risk. A composition theory would facilitate the estimation of the risk of re-identification of a linked data set using information from the source data.***- Data complexity is growing rapidly; the data is constantly being updated and growing. The resulting complex data sets will require big data de-identification methods to ensure they scale appropriately. There is a need for streaming de-identification methods that are designed specifically for health data sets.***- With the increasing availability of free-form medical text in EMRs, the analysis of this information is adding detail and context to structured data. More realistic evaluation frameworks for the de-identification of free-form text need to be developed that are designed specifically for the de-identification context, and then tools evaluated using that framework.***Our lab has been effective in transitioning its research results into practice through standards and software that have been adopted globally. We will continue this trend as we develop new methods from this research, facilitating the sharing of electronic health information for secondary purposes while protecting the privacy of patients and the identity of providers.********