De-Identification Challenges in Real-World Portuguese Clinical Texts
摘要
The electronic health record (EHR) data, widely used by hospitals and healthcare professionals, contain valuable information about the patient and treatments and has become increasingly relevant to clinical natural language processing (NLP) tasks. Although the growing number of EHR systems, these medical data contain sensitive information and cannot be released publicly without data security measures and privacy protection. The de-identification process is necessary to detect and remove all sensitive information in EHRs. There are several techniques to de-identify textual data, such as rule-based, machine learning and hybrid methods. This work proposes a new method to de-identify Portuguese clinical data based on dictionaries, phonetic normalization, similarity measures and regular expressions. We used actual EHR cardiology data from a Brazilian hospital to develop an algorithm to identify and replace data such as proper names, birth dates, credentials, emails, and phone numbers. Although still under development, our method correctly de-identified 318 documents (93.67% of the evaluated documents).