A practical method to improve the emergency department throughput

This is a 5-minute read discussing the core idea of my Master's thesis and provides an overview of my research without the technical methodologies.

Problem

First, let's dive into the problem that my thesis is trying to solve. As we all know, emergency department (ED) is the main point of entry for patients who require urgent care. Moreover, ED overcrowding is a well-known problem globally, affecting health care providers, their staff, and patients. Delays in the ED can have serious consequences since timely care is one of the most important aspects of the ED. The goal is to reduce the patient length of stay (LOS) in the ED by improving existing procedures.

Idea

Numerous factors contribute to ED overcrowding and one of them is the availability of diagnostic services such as radiology. More specifically I have focused on the prediction of computed tomography (CT) scans in the ED. A CT scan is a an advanced imaging technique that can create cross-sectional images, or slices, of the body. In many cases, CT scans require laboratory testing before the image acquisition can begin, which makes the patient preparation a time-consuming process.

Given that CT scans involve X-rays, only a physician can order a CT scan for the patient. However, by developing a model that can predict the need for a CT scan early in a patient's visit, we can complete the time-consuming preparation process while the patient is idle, awaiting the physician's initial assessment (PIA). The image at the beginning of the post illustrates the implementation of this approach.

Prediction model

For this approach to be effective, it is essential that the prediction can be done in the Arrival stage. This implies that the model input is limited to only the patient triage information. Triage is the process of categorizing patients based on the severity of their injuries and, by extension, the order in which patients receive care. This process is the first step after the arrival and usually do not take more than 5 minutes to complete. At this stage we have access to a small yet useful set of information about the patient and their issue. In my dataset I had access to patient's age, sex, mode of arrival (e.g., walk-in, ambulance, or other), triage acuity (a number from 1 to 5, 1 representing most ill and 5 least ill), chief complaint (a short text representing the patient's main reason to visit the ED), and treatment area (i.e., the area of the ED where the patient is assessed, monitored, and managed after the triage process).

This relatively small number of features makes the development of an accurate model very challenging. I experimented with various methods to identify the best-performing model. Given the limited number of features, more complex methods quickly led to overfitting of the dataset and performed poorly on the test set. After systematic experimentation with 210 prediction pipelines, the final method I used was a combination of a logistic regression model with SentenceBERT embeddings for the chief complaint. I also employed random oversampling to address the imbalanced classes in the dataset (only 13% of the ED patients require a CT scan).

Results and impact

To understand how this model can impact the ED we need to analyze the performance of the model. The model achieved an ROC AUC score of 0.86. But what does this actually mean? This metric is a good way to summarize the model performance in a single number. But I want to look a bit deeper on the model performance and see how it affects the ED in a few different scenarios. The Table below shows the sensitivity and specificity of the model. The former shows, out of 100 patients who require a CT scan, how many did the model recognize successfully and the latter shows, out of 100 patients who were predicted for a CT scan by the model, how many actually needed a CT scan. There is always a tradeoff between these two numbers and the best combination can be selected based on the specific scenario where the model is being deployed. The average LOS reduction is calculated for each model assuming that waiting for a CT scan after physician's order will take at most 30 minutes if the patient is already prepared at an earlier stage.

Model Specificity Sensitivity Average LOS reduction
1 80.0% 74.5% 41 minutes
2 90.0% 54.5% 30 minutes
3 95.0% 38.5% 21 minutes
4 99.0% 16.6% 9 minutes

This table shows that a significant reduction in patient length of stay can be achieved by implementing this model. In case of a false positive, no substantial burden is imposed on the patient. More interestingly, even if we aim for a very high specificity which translates to a lower sensitivity, the model is still beneficial because the patients who are predicted for a CT get a boosted preparation while others who were not detected would still go through the regular procedure.

Limitations

It is important to acknowledge the limitations of this study. Most notably, the data collected for this work was limited to a one-year period for a single hospital and may not be generalizable to other centers. Furthermore, although the proposed model can assist in predicting a CT scan during the triage process, it still does not account for all of the complexity in physician decision making.