Privacy preserving neural networks for electronic health records de-identification

Abstract

Over the last decade, significant improvements and efforts in digitizing healthcare provided us with a sizeable collection of electronic medical records. These Electronic Health Records (EHRs), especially the clinical narratives, brimming with hidden knowledge to be discovered, contain sensitive information about the patient. For this reason, medical institutions are legally prohibited from publishing these data in the raw format to the public. Hence, the recent surge towards finding a solution for de-identification detects and removes sensitive information from clinical narratives. Since 2016, we have seen several deep learning-based approaches for de-identification, which achieved over 98% accuracy. However, these models are trained with sensitive information and can unwittingly memorize some of its training data, and a careful analysis of these models can reveal patients' data. In this work, we propose a differentially private ensemble framework for de-identification, allowing medical researchers to collaborate through publicly publishing the de-identification models. To the best of our knowledge, this is the first privacy-preserving machine learning approach for the de-identification of EHRs. Experiments in three different datasets showed competitive results compared to the state-of-the-art methods with guaranteed differential privacy.

Ähnliche Arbeiten