Role of Local Context in Automatic Deidentification of Ungrammatical, Fragmented Text

Proceedings of the Human Language Technology Conference of the North American Chapter of the ACL, pages 65–73, New York, June 2006. c⃝2006 Association for Computational Linguistics Role of Local Context in Automatic Deidentification of Ungrammatical, Fragmented Text Tawanda Sibanda CSAIL Massachusetts Institute of Technology Cambridge, MA 02139 tawanda@mit.edu Ozlem Uzuner Department of Information Studies College of Computing and Information University at Albany, SUNY Albany, NY 12222 ouzuner@albany.edu Abstract Deidentification of clinical records is a crucial step before these records can be distributed to non-hospital researchers. Most approaches to deidentification rely heavily on dictionaries and heuristic rules; these approaches fail to remove most per- sonal health information (PHI) that cannot be found in dictionaries. They also can fail to remove PHI that is ambiguous between PHI and non-PHI. Named entity recognition (NER) tech- nologies can be used for deidentification. Some of these technologies exploit both local and global context of a word to iden- tify its entity type. When documents are grammatically written, global context can improve NER. In this paper, we show that we can dei- dentify medical discharge summaries us- ing support vector machines that rely on a statistical representation of local context. We compare our approach with three dif- ferent systems. Comparison with a rule- based approach shows that a statistical representation of local context contributes more to deidentification than dictionaries and hand-tailored heuristics. Compari- son with two well-known systems, SNoW and IdentiFinder, shows that when the lan- guage of documents is fragmented, local context contributes more to deidentifica- tion than global context. 1 Introduction Medical discharge summaries contain information that is useful to clinical researchers who study the interactions between, for example, different med- ications and diseases. However, these summaries include explicit personal health information (PHI) whose release would jeopardize privacy. In the United States, the Health Information Portability and Accountability Act (HIPAA) provides guide- lines for protecting the confidentiality of health care information. HIPAA lists seventeen pieces of textual PHI of which the following appear in medical dis- charge summaries: first and last names of patients, their health proxies, and family members; doctors’ first and last names; identification numbers; tele- phone, fax, and pager numbers; hospital names; ge- ographic locations; and dates. Removing PHI from medical documents is the goal of deidentification. This paper presents a method based on a statis- tical representation of local context for automati- cally removing explicit PHI from medical discharge summaries, despite the often ungrammatical, frag- mented, and ad hoc language of these documents, even when some words in the documents are am- biguous between PHI and non-PHI (e.g., “Hunting- ton” as the name of a person and as the name of a disease), and even when some of the PHI cannot be found in dictionaries (e.g., misspelled and/or for- eign names). This method differs from traditional approaches to deidentification in its independence from dictionaries and hand-tailored heuristics. It applies statistical named entity recognition (NER) methods to the more challenging task of deidenti- 65 fication but differs from traditional NER approaches …

Content truncated to 500 words