Use of Natural Language Processing to Identify 414 Different Chief Complaints in Adult Emergency Department Patients

 

Authors:

  • David A. Thompson 1
  • Mark Courtney 1
  • Sanjeev Malik 1
  • Michael J. Schmidt 1
  • Victoria Weston 2

1 Northwestern University, Northwestern University Feinberg School of Medicine

2 Northwestern Lake Forest Hospital

Background: Every medical encounter begins with a chief complaint, a reason for visit. Standard practice is to capture this as free-text, often in the patient’s own words.  The chief complaint can be an important data element for analytics of clinical workflow, quality improvement, public health, and education. However, for utility as data, it must be converted and codified into operational and standardized clinical concepts. Using an existing chief complaint vocabulary and natural language processing (NLP) engine, we codified and analyzed adult emergency department (ED) reason for visits.

Methods: All patients (86,948) presenting to an urban adult tertiary ED in 2016 were eligible for inclusion. Patients with missing data for age, gender, or reason for visit in the electronic health record were excluded, as were all patients under 18 years of age. We used a commercially available NLP engine (Health Navigator LLC) to convert the patient’s free-text reason for visit into a standardized clinical vocabulary of Coded Chief Complaints. For each chief complaint the frequency, admission rate, and average Emergency Severity Index (ESI) score were calculated.

Results: The NLP engine identified and coded 414 different chief complaints, for 42,532 different reason for visit text strings, from 80,579 adult ED patient encounters. The top 10 chief complaints were abdomen pain (9.6% of all encounters; 32% admitted; ESI average 2.7; average age 43 years), chest pain (8.6%; 46%; 2.2; 49), shortness of breath (3.8%; 60%; 2.8; 56), back pain (3.6%; 20%; 3.1; 47), vomiting (3.3%; 40%; 2.6; 44), alcohol use and abuse (3.3%; 10%; 2.6; 41), head pain (3.2%; 18%; 2.7; 43), dizziness (3.2%; 37%; 2.4; 52), fever (2.6%; 55%; 2.5; 48), and laceration (2.0%; 4%; 3.5; 42). There were 29,688 patients admitted to the hospital. The top 10 chief complaints that resulted in admission were chest pain (10.7% of all admissions), abdomen pain (8.4%), shortness of breath (6.2%), fever (3.9%), vomiting (3.6%), dizziness (3.2%), weakness – generalized (2.1%), loss of consciousness (2.1%), back pain (2.0%), and altered mental status (1.6%).

Conclusion: NLP can be used to consolidate the free text heterogeneity of over 80,000 adult ED encounters into just over 400 discrete reasons for visit. This approach provides a methodology and new data source for future emergency department research.

Chart shows a histogram sorted left to right by the frequency of the top 25 Coded Chief Complaints found in this adult emergency department patient population. Heath Navigator also provides an acuity rating for each Code Chief Complaint. This can be visualized as a heat map from dark red to dark green.

Citation:  Thompson DA, Courtney DM, Malik S, Schmidt MJ, Weston V. Use of Natural Language Processing to Identify 414 Different Chief Complaints in Adult Emergency Department Patients. Acad Emerg Med. 2018;25 (Issue S1):S139.

Related References

  • Thompson DA, Eitel D. Coded Chief Complaints – Automated Analysis of Free-Text Complaints. Fernandes CMB, Pines JM, Amsterdam J, Davidson SJ. Acad Emerg Med. 2006;13:774-782. PubMed Abstract. Abstract and Chart.

Related Links

View Research and References.