Leveraging Task Transferability to Meta-learning for Clinical Section Classification with Limited Data

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 6690 - 6702 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Leveraging Task Transferability to Meta-learning for Clinical Section Classification with Limited Data Zhuohao Chen1∗, Jangwon Kim2, Ram Bhakta2†, Mustafa Y. Sir2 1University of Southern California, Los Angeles, USA 2Amazon, Seattle, USA zhuohaoc@usc.edu {jangwok, simustaf}@amazon.com hiten.bhakta@gmail.com Abstract Identifying sections is one of the critical com- ponents of understanding medical information from unstructured clinical notes and devel- oping assistive technologies for clinical note- writing tasks. Most state-of-the-art text clas- sification systems require thousands of in- domain text data to achieve high performance. However, collecting in-domain and recent clin- ical note data with section labels is challenging given the high level of privacy and sensitivity. This paper proposes an algorithmic way to im- prove the task transferability of meta-learning- based text classification in order to address the issue of low-resource target data. Specifically, we explore how to make the best use of the source dataset and propose a unique task trans- ferability measure named Normalized Nega- tive Conditional Entropy (NNCE). Leveraging the NNCE, we develop strategies for selecting clinical categories and sections from source task data to boost cross-domain meta-learning accuracy. Experimental results show that our task selection strategies improve section clas- sification accuracy significantly compared to meta-learning algorithms. 1 Introduction An important part of Electronic Health Records (EHRs) is the digitized clinical notes that contain the medical and treatment histories of patients. The section of clinical notes can be defined as a text seg- ment that clusters consecutive sentences with rele- vant content of one dimension of a patient’s health encounter (Pomares-Quimbaya et al., 2019). Clin- ical note sections, labeled with either headings or subheadings, make the notes well organized and of- fer improved clinical information extraction (Wang et al., 2018b). However, many clinical notes con- tain narratives that are in an unstructured free-text ∗Work done during an internship at Amazon Care †Work done while Ram Bhakta was a researcher at Ama- zon Care in 2021, and Ram is now affiliated to Oath Care, Austin, USA format, (e.g., History of Present Illnesses described in paragraph form), which makes it challenging to retrieve and utilize this information. In the United States, physicians generally spend an excessive amount of time interfacing with EHRs and com- puterized physician order entry (CPOE) workflows in their aftercare work, resulting in burnout, low job satisfaction, and system-wise inefficiencies (Pa- tel et al., 2018). An automated section classifier can play a key role in mitigating this problem. In some cases, section classification serves as an end task of automatic report segmentation. For exam- ple, according to an internal survey we conducted with Amazon Care providers, we found evidence that classifying sentences related to the History of Present Illness from medical encounters can greatly assist providers with their documentation. For computer-assisted report generation, understand- ing clinical notes from an unstructured format is an important data pre-processing (Gopinath et al., 2020). There are …

Content truncated to 500 words