Extracting Symptoms and their Status from Clinical Conversations

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 915–925 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 915 Extracting Symptoms and their Status from Clinical Conversations Nan Du∗, Kai Chen, Anjuli Kannan, Linh Tran, Yuhui Chen, and Izhak Shafran {dunan, kaichen, anjuli, tranlm, yuhuic, izhak}@google.com Google Inc. Abstract This paper describes novel models tailored for a new application, that of extracting the symptoms mentioned in clinical conversations along with their status. Lack of any publicly available corpus in this privacy-sensitive do- main led us to develop our own corpus, con- sisting of about 3K conversations annotated by professional medical scribes. We propose two novel deep learning approaches to infer the symptom names and their status: (1) a new hierarchical span-attribute tagging (SA-T) model, trained using curriculum learning, and (2) a variant of sequence-to-sequence model which decodes the symptoms and their sta- tus from a few speaker turns within a slid- ing window over the conversation. This task stems from a realistic application of assist- ing medical providers in capturing symptoms mentioned by patients from their clinical con- versations. To reflect this application, we de- fine multiple metrics. From inter-rater agree- ment, we find that the task is inherently diffi- cult. We conduct comprehensive evaluations on several contrasting conditions and observe that the performance of the models range from an F-score of 0.5 to 0.8 depending on the con- dition. Our analysis not only reveals the in- herent challenges of the task, but also provides useful directions to improve the models. 1 Introduction In recent years, hospitals and clinics across the United States have been coaxed and cajoled into adopting Electronic Health Records through pub- lic policies and insurance requirements. This has led to the unforeseen side-effect of placing a dis- proportionate burden of documentation on physi- cians, causing burnouts among them (Wachter and Goldsmith, 2018; Xu, 2018). One study found that full-time primary care physicians spent about 4.5 ∗All the authors contributed equally. hours of an 11-hour workday interacting with the clinical documentation systems, and yet were still unable to finish their documentations and had to spend an additional 1.4 hours after normal clinical hours (Arndt et al., 2017). Speech and natural language processing are now sufficiently mature that there has been con- siderable interest, both in academia and industry, to investigate how these technologies can be ex- ploited to simplify the task of documentation, and to allow physicians to dedicate more time to pa- tients. While domain-specific ASR systems that allow doctors to dictate notes have been around for a while, recent work (Patel et al., 2018; Finley et al., 2018a,b) has begun to address more chal- lenging tasks, such as extracting relevant informa- tion directly from doctor-patient conversations. In this work, we investigated the task of in- ferring symptoms mentioned in clinical conver- sations, along with whether patients have experi- enced them or not. Our contributions include: (i) defining the task, including the annotation scheme for labeling the …

Content truncated to 500 words