B0599
Title: Preserving patient privacy in dynamic treatment regimes: Private outcome-weighted learning (PrOWL)
Authors: Dylan Spicker - University of New Brunswick (Saint John) (Canada) [presenting]
Erica Moodie - McGill University (Canada)
Susan Shortreed - (United States)
Abstract: Precision medicine is a branch of evidence-based medicine which leverages individual-level characteristics to inform treatment recommendations. Dynamic treatment regimes (DTRs) are one framework for formalizing precision medicine. Because of the sensitive nature of medical data, it is critical that researchers working with DTRs guarantee safety and privacy for study participants. There is a growing literature that demonstrates that summary statistics and model parameters may leak private information about the individuals in a dataset. These findings have sparked interest in the study of differential privacy: a formal standard of privacy which provides rigorous guarantees of the protection afforded to individuals. A new method is presented for DTR estimation, called private outcome-weighted learning (PrOWL), based on the existing outcome-weighted learning (OWL) techniques for DTR estimation. PrOWL achieves differential privacy and has provable accuracy bounds. PrOWL allows for the application of these familiar and effective OWL-based procedures while providing treatment rules which are private and can be publicly released. In addition to the theoretical accuracy bounds, PrOWL is explored through simulation, demonstrating the strengths and limitations of the method. Beyond the proposal of PrOWL, the importance of considering privacy in precision medicine is highlighted and important areas are illustrated for future investigation.