Machine learning for selecting patients with Crohn's disease for abdominopelvic computed tomography in the emergency department

Background: Patients with Crohn's disease (CD) frequently undergo abdominopelvic computed tomography (APCT) in the emergency department (ED). It's essential to diagnose clinically actionable findings (CAF) as they may need immediate intervention, frequently surgical. However, repeated APCT's includes increased ionizing radiation exposure. Guidance regarding APCT performance is mostly clinical and empiric. Aims: We used a machine learning (ML) approach for predicting CAF on APCT in the ED. Methods: We performed a retrospective cohort study of patients with CD who presented to the ED and underwent APCT. CAF were defined as bowel obstruction, perforation, intra-abdominal abscess or complicated fistula. ML was used to predict the probability of having CAF on APCT, using routine clinical variables. Results: Of 101 admissions included, 44 (43.5%) had CAF on APCT. ML successfully identified patients at low (NPV 91.6%, CI-95% 90.6–92.5) and high (PPV 92.8%, CI-95%, 92.3–93.2) risk for CAF (AUROC = 0.774), using beats-per-minute, mean arterial pressure, neutrophil-to-lymphocyte ratio and sex. This allowed the construction of a risk stratification scheme according to patients’ probability for CAF on APCT. Conclusion: We present a novel artificial intelligence-based approach, utilizing readily available clinical variables to better select patients with CD in the ED for APCT. This might reduce the number of APCTs performed, avoiding related hazards while ensuring high-risk patients undergo APCT.