Closing the Gap: Joint De-Identification and Concept Extraction in the Clinical Domain

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6945–6952 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 6945 Closing the Gap: Joint De-Identification and Concept Extraction in the Clinical Domain Lukas Lange1,2,3 Heike Adel1 1 Bosch Center for Artificial Intelligence, Renningen, Germany 2 Spoken Language Systems (LSV), Saarland University, Saarbr¨ucken, Germany 3 Saarbr¨ucken Graduate School of Computer Science, Saarbr¨ucken, Germany {Lukas.Lange,Heike.Adel,Jannik.Stroetgen}@de.bosch.com Jannik Str¨otgen1 Abstract Exploiting natural language processing in the clinical domain requires de-identification, i.e., anonymization of personal information in texts. However, current research considers de-identification and downstream tasks, such as concept extraction, only in isolation and does not study the effects of de-identification on other tasks. In this paper, we close this gap by reporting concept extraction perfor- mance on automatically anonymized data and investigating joint models for de-identification and concept extraction. In particular, we pro- pose a stacked model with restricted access to privacy-sensitive information and a multi- task model. We set the new state of the art on benchmark datasets in English (96.1% F1 for de-identification and 88.9% F1 for concept extraction) and Spanish (91.4% F1 for concept extraction). 1 Introduction In the clinical or biomedical domain, natural lan- guage processing (NLP) could significantly im- prove the efficiency and effectiveness of pro- cesses. For example, the extraction of struc- tured information from clinical narratives can help in decision making or drug repurposing (Mari- mon et al., 2019). However, the automatic pro- cessing of documents with privacy-sensitive con- tent is restricted due to the necessity of applying anonymization techniques. Text anonymization, also called de- identification, aims at detecting and replacing protected health information (PHI),1 such as patient names, age and phone numbers, as shown in the upper part of Figure 1. Recent studies show that automatic de-identification leads to 1PHI types are typically defined by governments, for in- stance in the Health Insurance Portability and Accountability Act (HIPAA) of the United States. PATIENT AGE PROBLEM Patrick died of myocardial infarction (MI) at 76.  PHI terms  concept  Figure 1: Sentence with annotations of the two tasks. promising results (Stubbs and Uzuner, 2015; Marimon et al., 2019). A severe limitation of current approaches in research, however, is that de-identification is typically addressed in isolation but not together with a downstream task, such as concept extraction (CE) from medical texts (Gonzalez-Agirre et al., 2019; Uzuner et al., 2011). Instead, the downstream task models are trained and evaluated on the non-anonymized data, and it remains unclear how de-identification affects their performance in real-world settings. In this paper, we argue that to evaluate the ef- fectiveness of NLP in the medical domain, the tasks of de-identification and information extrac- tion should be analyzed together. Our contribu- tions are as follows: We close this gap and ana- lyze the effect of de-identification on clinical con- cept extraction. Moreover, we consider the two tasks jointly and propose two end-to-end models: A multitask model that shares the input represen- tation across tasks, and a stacked model …

Content truncated to 500 words