A Semi-supervised Approach for De-identification of Swedish Clinical Text

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 4444–4450 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 4444 A Semi-supervised Approach for De-identification of Swedish Clinical Text Hanna Berg, Hercules Dalianis Department of Computer and Systems Sciences Stockholm University {hanna.berg, hercules}@dsv.su.se Abstract An abundance of electronic health records (EHR) is produced every day within healthcare. The records possess valuable information for research and future improvement of healthcare. Multiple efforts have been done to protect the integrity of patients while making electronic health records usable for research by removing personally identifiable information in patient records. Supervised machine learning approaches for de-identification of EHRs need annotated data for training, annotations that are costly in time and human resources. The annotation costs for clinical text is even more costly as the process must be carried out in a protected environment with a limited number of annotators who must have signed confidentiality agreements. In this paper is therefore, a semi-supervised method proposed, for automatically creating high-quality training data. The study shows that the method can be used to improve recall from 84.75% to 89.20% without sacrificing precision to the same extent, dropping from 95.73% to 94.20%. The model’s recall is arguably more important for de-identification than precision. Keywords: semi-supervised learning, self training, de-identification, Swedish clinical text 1. Introduction An abundance of electronic health records is produced ev- ery day. They possess valuable information for research and future improvement of medical care. A great majority will, however, never be re-used for research. This is partly due to the presence of sensitive data. Sensitive data may re- veal the identity of individual patients. Sensitive data also restricts the possibility to re-use the valuable information in electronic health records for research. Sensitive data must, therefore, be removed to protect the integrity of patients. Multiple efforts have been done to protect the integrity of patients while making electronic health records usable for research by removing personally identifiable information in patient records (Velupillai et al., 2009; Stubbs et al., 2017; Dernoncourt et al., 2017). Personally identifiable in- formation may, for example, be personal names, addresses, phone numbers, dates and locations. Standard methods for de-identification is Named Entity Recognition by identify- ing personally identifiable information and then removing them (Meystre, 2015). The American HIPAA Privacy Rule states that a health record is de-identified if eighteen types of identifiers are removed from the record (HIPAA, 2003). De-identification systems are designed to identify and remove this personally identifiable information (Velupillai et al., 2009; Stubbs et al., 2015). The annotation process for supervised machine learning methods is costly (Haertel et al., 2008). The annotation costs for clinical text is even more costly as the process must be carried out in a protected environment with a limited number of annotators who must have signed confidentiality agreements. Dernoncourt et al. (2017) estimated, from numbers presented in (Neamatullah et al., 2008)1, that manual de-identification of the clinic MIMIC dataset, would require at least two annotators and …

Content truncated to 500 words