ARRS 2022 Abstracts

RETURN TO ABSTRACT LISTING


1987. Advanced AI Classification Methods Can Select the Most Accurate and Timely Modality to Diagnose Appendicitis in Children
Authors * Denotes Presenting Author
  1. Edward Florez *; University of Mississippi Medical Center
  2. Jessie Smith; University of Mississippi Medical Center
  3. Candace Howard; University of Mississippi Medical Center
Objective:
According to national recommendations, ultrasound (US) is the first-line imaging modality for diagnosing pediatric abdominal pain in emergency room visitors, often confirmed as appendicitis. However, the clinical diagnosis of acute appendicitis in children can be challenging due to limitations in the physical examination, mainly in overweight or obese pediatric patients. Thus, the use of radiation-based imaging such as computed tomography (CT) is performed after US but still with many safety concerns. This study will attempt to select the most accurate and timely modality (US vs CT) for this illness based on artificial intelligence (AI) classification models to avoid additional imaging assessment, which may potentially delay management and increased cost and risk of complications in the initial evaluation of suspected pediatric appendicitis.

Materials and Methods:
This is an IRB-approval retrospective study of 1111 pediatric patients with a history of appendicitis-like symptoms admitted to the emergency room between 2015 and 2019. Patients who underwent both CT and US were included in the study (N=396, 203 girls and 193 boys, 9.11 +/- 4.14 years old at US and/or CT scan). Demographic and clinical information such as age, BMI, gender, and Alvarado score were collected from the electronic medical record. Additionally, anthropometric measurements such as waist circumference (WC) and the sagittal abdominal diameter (SAD) were measured. Different approaches were implemented using in-house algorithms written in Python to select the best modality to diagnose appendicitis using two different types of analyses: (1) bivariate analysis which involved two independent variables such as BMI and SAD; and (2) multivariate analysis, which included parameters such as age, gender, BMI, PMNs(%), ANC, Alvarado Score, WC, and, SAD. The models tested were logistic regression (LR), K-nearest neighbors (K-NN), support vector machine (SVM), Naive Bayes (NB), decision tree (DT), and random forest (RF). The full cohort was split into training and test sets before running the algorithms. Diverse training set-to-test set ratios were tested and compared: 80/20, 85/15, 90/10, and 95/05. Statistical analysis was performed using IBM SPSS statistics software version 25. Comparisons between the models were based on their prediction accuracy and standard deviation.

Results:
Our previous predictor model was based on BMI interactions (AUC=0.65, p<0.001). In this study, all AI classification models (except NB) had good predictive accuracy (>0.70) in the bivariate analysis when used train/test ratios equal to 0.90/0.10 and 0.95/0.05. In addition, DT and RF had a strong predictive accuracy (>0.80) in the multivariate analysis for train/test ratio equal to 0.95/0.05.

Conclusion:
This study shows that advanced AI classification methods can select the most accurate and timely modality to diagnose appendicitis in children, thereby avoiding additional imaging assessment that may potentially delay management and increasing cost and risk of complications in a diverse pediatric population.