Research Insights

Business Statistics

Topic Modeling on Triage Notes with Semi-orthogonal Non-negative Matrix Factorization

Yutong Li, Ruoqing Zhu, Annie Qu, Han Ye, Zhankun Sun
Published in Journal of the American Statistical Association, December 2020

Emergency Department (ED) overcrowding is a universal health issue that impairs the public access to emergency care. The medical society has long understood the root cause of ED overcrowding, i.e., the prolonged occupation of ED beds by patients who are admitted and waiting to be transferred to inpatient beds. This phenomenon is also referred to as ED blocking. One initiative to alleviate ED blocking is to put in a request for inpatient beds at triage for patients who are very likely to be admitted, even before the patients are seen by physicians. Inpatient beds are scarce and expensive resources. Hence, such decisions are contingent on the prediction accuracy of the admission decisions.

In this case study, Dr Zhankun Sun, Assistant Professor in the Department of Management Sciences and co-authors, build a classifier to predict the disposition of patients using manually typed nurse notes collected during triage in a large teaching hospital.

"Our data analysis on 600,000 electronic health records shows that the triage notes contain strong predictive information towards classifying the disposition of patients for certain medical complaints, such as altered consciousness or stroke," says Sun.

This improvement could be clinically impactful for certain patients, especially when the scale of hospital patients is large. Furthermore, the generated word-topic vectors provide a bi-clustering interpretation under each topic due to the orthogonal formulation, which can be beneficial for hospitals in better understanding the symptoms and reasons behind patients' visits.

These predictions can potentially be incorporated to early bed coordination and fast track streaming strategies to reduce waiting times and alleviate overcrowding in the ED. However, these triage notes involve high dimensional, noisy, and sparse text data, which make model-fitting and interpretation difficult. To address this issue, they propose a novel semi-orthogonal non-negative matrix factorization for both continuous and binary predictors to reduce the dimensionality and derive word topics. The triage notes can then be interpreted as a non-subtractive linear combination of orthogonal basis topic vectors.