A. M. Polcari1, A. J. Benjamin1, A. Pratt2, K. Wilson1 1University Of Chicago, Department Of Surgery, Chicago, IL, USA 2University Of Chicago, Department Of Emergency Medicine, Chicago, IL, USA
Introduction: The top reason for Emergency Department (ED) visits in the United States is abdominal pain. More than 3 million patients per year are admitted urgently or emergently for an Emergency General Surgery (EGS) diagnosis, 25% of which undergo an operation. Swift ED throughput – time from triage to disposition – for EGS patients is dependent on early recognition of an acute problem, rapid laboratory and imaging study completion, and prompt surgical consult. A national crisis of ED crowding has led to bottlenecks in throughput, increasing complications. Currently, designation of patient acuity is done solely by ED triage nurses using the Emergency Severity Index (ESI), a largely subjective measure. We sought to create an Abdominal Pain Triage (APT) score to rapidly identify ED patients requiring urgent operative intervention via machine learning.
Methods: Demographic, vital signs, laboratory data, ED triage-assigned ESI, and time to operative intervention were abstracted for all patients presenting to our institution’s adult ED from January 2011 to October 2021 with a chief complaint of abdominal pain, as identified using ICD-9 and ICD-10 codes. XGBoost, a decision-tree-based machine learning algorithm, was used to construct a model predicting operative intervention within 12 hours of presentation to the ED. The dataset was split into training (80%) and test (20%) sets. AU-ROC was calculated to compare the APT model performance relative to triage-nurse designated ESI.
Results: A total of 50,269 patients were identified with a chief of abdominal pain, 2,052 of which had an operation within 12 hours of presentation. The APT model had strong predictive power for patients requiring an operation (AU-ROC 0.80). Feature importance revealed white blood cell count to be the greatest predictor of undergoing operative intervention. Compared to triage nurse-assigned ESI scores (AUROC 0.67), the APT score had significantly greater discriminative ability in identifying ED patients requiring urgent surgical intervention (Figure 1).
Conclusion: To our knowledge, this is the first machine learning algorithm developed to identify EGS patients requiring urgent surgical intervention based on initial triage and lab data. The APT score accurately differentiates patients who present to the ED with abdominal pain requiring urgent surgical intervention from those who do not. The APT score can be implemented in real time for dynamic triage of patients presenting to the ED with abdominal pain, thus expediting time to surgery and improving EGS outcomes.
