Skip to Content
SBMI logo

An evaluation of an automated system for estimation of emergency severity index (ESI) using a probabilistic graphical model

Author: Chaitanya Churi

Primary Advisor: Dean Sittig, PhD

Committee Members: Jamison Feramisco, Nicholas Peterson, Neal Sikka, Kim Branson

Masters thesis, The University of Texas Health Science Center School of Health Information Sciences at Houston.

Abstract:

The Emergency Severity Index (ESI) is a 5-point triage scale used by emergency departments (EDs) to prioritize patients who present with varying complaints. The present study aimed to assess the accuracy of the Lumiata automated triage system that predicts ESI scores based on a probabilistic graphical model of medicine. Patient records of 74,549 consecutive ED visits were automatically analyzed and a triage score was predicted for each. These predictions were compared to the original ESI ratings, along with predictions from random null-predictor and logistic regression models. A subset of these ratings was also compared to physician-rated triage scores. Percent agreement, error rates and linear-weighted kappa values were calculated for each predictor. The automated system showed 40.3%-42.2% agreement with the original data and performed better than the null predictor and half the physician raters. The logistic regression model demonstrated higher agreement and kappa values, which is attributable to the higher frequencies of its predicted ESI level 3. This model performed poorly when compared to the Lumiata system across all other ESI levels, as noted by higher error rates. Given the subjective nature of ESI prediction and limited to only medical record data, the Lumiata system demonstrated significant information gain and has potential applications in emergency triage.