A Recurrent Neural Network Architecture for De-identifying Clinical Records

Shweta,A. Kumar,Asif Ekbal,S. Saha,P. Bhattacharyya

2016-12-01

Abstract:Electronic Medical Records contains a rich source of information for medical finding. However, the access to the medical record is limited to only de-identified form so as to protect the confidentiality of patient. According to Health Insurance Portability and Accountability Act, there are 18 PHI categories that should be enclosed before making the EMR publicly available. With the rapid growth of EMR and a limited amount of de-identified text, the manual curation is quite unfeasible and time-consuming, which has drawn the attention of several researchers to propose automated de-identification system. In this paper, we proposed deep neural network based architecture for de-identification of 7 PHI categories with 25 associated sub-categories. We used standard benchmark dataset from i2b2-2014 de-identification challenge and performed the comparison with very strong baseline based on Conditional Random Field. We also perform the comparison with the state-of-art. Results show that our proposed system achieves significant improvement over baseline and comparable performance over state-of-art.