Automated de-identification of free-text medical records

Abstract

This paper presents a de-identification study at the Harvard-MIT Division of Health Science and Technology (HST) to automatically de-identify confidential patient information from text medical records used in intensive care units (ICUs). Patient records are a vital resource in medical research. Before such records can be made available for research studies, protected health information (PHI) must be thoroughly scrubbed according to HIPAA specifications to preserve patient confidentiality. Manual de-identification on large databases tends to be prohibitively expensive, time-consuming and prone to error, making a computerized algorithm an urgent need for large-scale de-identification purposes. We have developed an automated pattern-matching deidentification algorithm that uses medical and hospital-specific information. The current version of the algorithm has an overall sensitivity of around 0.87 and an approximate positive predictive value of 0.63. In terms of sensitivity, it performs significantly better than 1 person (0.81) but not quite as well as a consensus of 2 human de-identifiers (0.94). The algorithm will be published as open-source software, and the de-identified medical records will be incorporated into HST's Multi-Parameter Intelligent Monitoring for Intensive Care (MIMIC II) physiologic database.