RESUMO
INTRODUCTION: There is a tendency toward nonoperative management of appendicitis resulting in an increasing need for preoperative diagnosis and classification. For medical purposes, simple conceptual decision-making models that can learn are widely used. Decision trees are reliable and effective techniques which provide high classification accuracy. We tested if we could detect appendicitis and differentiate uncomplicated from complicated cases using machine learning algorithms. MATERIALS AND METHODS: We analyzed all cases admitted between 2010 and 2016 that fell into the following categories: healthy controls (Group 1); sham controls (Group 2); sham disease (Group 3), and acute abdomen (Group 4). The latter group was further divided into four groups: false laparotomy; uncomplicated appendicitis; complicated appendicitis without abscess, and complicated appendicitis with abscess. Patients with comorbidities and whose complete blood count and/or pathology results were lacking were excluded. Data were collected for demographics, preoperative blood analysis, and postoperative diagnosis. Various machine learning algorithms were applied to detect appendicitis patients. RESULTS: There were 7244 patients with a mean age of 6.84 ± 5.31 years, of whom 82.3% (5960/7244) were male. Most algorithms tested, especially linear methods, provided similar performance measures. We preferred the decision tree model due to its easier interpretability. With this algorithm, we detected appendicitis patients with 93.97% area under the curve (AUC), 94.69% accuracy, 93.55% sensitivity, and 96.55% specificity, and uncomplicated appendicitis with 79.47% AUC, 70.83% accuracy, 66.81% sensitivity, and 81.88% specificity. CONCLUSIONS: Machine learning is a novel approach to prevent unnecessary operations and decrease the burden of appendicitis both for patients and health systems. LEVELS OF EVIDENCE: III.