A New Public Corpus for Clinical Section Identification: MedSecId

Proceedings of the 29th International Conference on Computational Linguistics, pages 3709–3721 October 12–17, 2022. 3709 A New Public Corpus for Clinical Section Identification: MedSecId Paul Landes:, Kunal Patel;, Sean S. Huang˛, Adam Webb;, Barbara Di Eugenio:, and Cornelia Caragea: :Department of Computer Science, University of Illinois at Chicago ;Department of Emergency Medicine, University of Illinois at Chicago ˛Department of Internal Medicine and Geriatrics, University of Illinois at Chicago {plande2,kpate318,sshuang,awebb25,bdieugen,cornelia}@uic.edu Abstract The process by which sections in a document are demarcated and labeled is known as section identification. Such sections are helpful to the reader when searching for information and con- textualizing specific topics. The goal of this work is to segment the sections of clinical med- ical domain documentation. The primary con- tribution of this work is MedSecId, a publicly available set of 2,002 fully annotated medical notes from the MIMIC-III. We include several baselines, source code, a pretrained model and analysis of the data showing a relationship be- tween medical concepts across sections using principal component analysis. 1 Introduction Most unstructured medical text found in electronic health record systems (EHRs) written by medical staff have conceptually well defined sections. For example, discharge summaries are technical med- ical documents, written by physicians when the pa- tient is discharged, which describe the patient’s hos- pital stay and surrounding circumstances of their illness. As shown by the example in Figure 1, dis- charge summaries consist of named sections, typ- ically in a specific sequence, such as the History of Present Illness; this type of section appears both in discharge summaries and in physician notes that describe a chronology of an illness that begins with the admission of the patient. Whereas sections often have headers, section identification (SI) is more challenging than simply parsing the first several leading header tokens of the respective section (underlined in Figure 1). While the first several tokens can be helpful in identifying a section, their naming often varies. For exam- ple, the 6th section in Figure 1 starts with header tokens Preoperative Laboratory Data, but the sec- tion type is labs-imaging. There are also cases where the header tokens are missing, as shown in Admission Date: [**2126-2-7**] Discharge Date: [**2126-2-20**] Date of Birth: [**2069-4-1**] Sex: M history-of-present-illness HISTORY OF PRESENT ILLNESS: Mr. [**Known last- name **] is a 56-year-old male who experienced chest. .. past-medical-history PAST MEDICAL HISTORY: Hypertension, former smoker with a 4- pack per day history for which he. . . social-history SOCIAL HISTORY: He lives alone, and he works at [**Hospital3 2576**] as a cargo transporter. medication-history MEDICATIONS ON ADMISSION: Aspirin 325 mg p.o. once a day, Toprol-XL 50 mg p.o. once a day. allergies ALLERGIES: He had no known drug allergies. labs-imaging PREOPERATIVE LABORATORY DATA: White count 6.0, hematocrit 33.3, platelet count 329,000.. . . hospital-course On exam he had a left facial droop, status post his child- hood polio. Temperature of 97.5, heart rate 65 in sinus. . . discharge-diagnosis DISCHARGE DIAGNOSES: 1. Status post coronary artery bypass grafting x 3.. . …

Content truncated to 500 words