The following is a summary of the “Machine Learning Methods for Predicting Patient-Level Emergency Department Workload,” published in the January 2023 issue of Emergency Medicine by Joseph et al.

Work Relative Value Units (wRVUs) are incorporated into various salary structures as a measure of the time and effort put into patient care. Therefore, predicting the number of wRVUs a patient will generate at triage with high accuracy would have many operational and clinical benefits, including easing the burden on individual doctors by spreading their workload more evenly. This study used data typically collected during triage to test whether deep-learning methods could accurately predict a patient’s wRVUs. Participants were adults who visited an urban, academic ER between July 1, 2016, and March 1, 2020. 

Structured data (age, sex, vital signs, Emergency Severity Index score, language, race, standardized chief complaint) and unstructured data (free-text chief complaint) were used in the de-identified triage information, with wRVUs serving as the outcome measure. The researchers looked at five models, including the mean wRVUs per the primary complaint, linear regression, neural networks, gradient-boosted trees on structured data, and neural networks on unstructured textual data. A mean absolute error was used as a metric to rank the quality of the models. Between January 1, 2016, and February 28, 2020, they analyzed 204,064 visits. Age, gender, and race significantly affected wRVUs, with the median wRVUs being 3.80 (interquartile range 2.56-4.21). 

Model errors decreased with increasing model complexity. Predictions with the use of chief complaints demonstrated a mean error of 2.17 wRVUs per visit (95% CI 2.07-2.27). The linear regression model showed an error of 1.00 wRVUs (95% CI 0.97-1.04), the gradient-boosted tree showed an error of 0.85 wRVUs (95% CI 0.84-0.86), the neural network with structured data showed an error of 0.86 wRVUs (95% CI 0.85-0.87), and the neural network with unstructured data showed an error of 0.78 wRVUs (95% CI 0.76-0.80). However, deep learning techniques show promise in overcoming the limitations of chief complaints as a predictor of the time required to evaluate a patient. These algorithms may have numerous useful applications, such as reducing bias in the triage process, quantifying crowding and mobilizing resources, and balancing emergency physicians’ and compensation.

Source: sciencedirect.com/science/article/abs/pii/S0736467922005686