Application of Artificial Neural Network Models to Differentiate Between Complicated and Uncomplicated Acute Appendicitis

Preoperative prediction of complicated appendicitis is challenging, and many clinical tools are developed to predict complicated appendicitis. This study evaluated whether a supervised learning method can recognize complicated appendicitis in emergency department (ED). Consecutive patients with acute appendicitis presenting to the ED were enrolled and included into training and testing datasets at a ratio of 70:30. The multilayer perceptron artificial neural network (ANN) models were trained to perform binary outcome classification between uncomplicated and complicated acute appendicitis. Measures of sensitivity, specificity, positive and negative likelihood ratio (LR + and LR-), and a c statistic of a receiver of operating characteristic curve were used to evaluate an ANN model. The simplest ANN model by Bröker et al. including the C-reactive protein (CRP) and symptom duration as variables achieved a c statistic value of 0.894. The ANN models developed by Avanesov et al. including symptom duration, appendiceal diameter, periappendiceal fluid, extraluminal air, and abscess as variables attained a high diagnostic performance (a c statistic value of 0.949) and good efficiency (sensitivity of 78.6%, specificity of 94.5%, LR + of 14.29, LR- of 0.23 in the testing dataset); and our own model by H.A. Lin et al. including the CRP level, neutrophil-to-lymphocyte ratio, fat-stranding sign, appendicolith, and ascites exhibited high accuracy (c statistic of 0.950) and outstanding efficiency (sensitivity of 85.7%, specificity of 91.7%, LR + of 10.36, LR- of 0.16 in the testing dataset). The ANN models developed by Avanesov et al. and H.A. Lin et al. developed model exhibited a high diagnostic performance.