Diagnosis and classification of pediatric acute appendicitis by artificial intelligence methods: An investigator-independent approach.

• Main Authors: Josephine Reismann, Alessandro Romualdi, Natalie Kiss, Maximiliane I Minderjahn, Jim Kallarackal, Martina Schad, Marc Reismann
• Format: Article
• Language: English
• Published: Public Library of Science (PLoS) 2019-01-01
• Series: PLoS ONE
• Online Access: https://doi.org/10.1371/journal.pone.0222030

Acute appendicitis is one of the major causes for emergency surgery in childhood and adolescence. Appendectomy is still the therapy of choice, but conservative strategies are increasingly being studied for uncomplicated inflammation. Diagnosis of acute appendicitis remains challenging, especially due to the frequently unspecific clinical picture. Inflammatory blood markers and imaging methods like ultrasound are limited as they have to be interpreted by experts and still do not offer sufficient diagnostic certainty. This study presents a method for automatic diagnosis of appendicitis as well as the differentiation between complicated and uncomplicated inflammation using values/parameters which are routinely and unbiasedly obtained for each patient with suspected appendicitis. We analyzed full blood counts, c-reactive protein (CRP) and appendiceal diameters in ultrasound investigations corresponding to children and adolescents aged 0-17 years from a hospital based population in Berlin, Germany. A total of 590 patients (473 patients with appendicitis in histopathology and 117 with negative histopathological findings) were analyzed retrospectively with modern algorithms from machine learning (ML) and artificial intelligence (AI). The discovery of informative parameters (biomarker signatures) and training of the classification model were done with a maximum of 35% of the patients. The remaining minimum 65% of patients were used for validation. At clinical relevant cut-off points the accuracy of the biomarker signature for diagnosis of appendicitis was 90% (93% sensitivity, 67% specificity), while the accuracy to correctly identify complicated inflammation was 51% (95% sensitivity, 33% specificity) on validation data. Such a test would be capable to prevent two out of three patients without appendicitis from useless surgery as well as one out of three patients with uncomplicated appendicitis. The presented method has the potential to change today's therapeutic approach for appendicitis and demonstrates the capability of algorithms from AI and ML to significantly improve diagnostics even based on routine diagnostic parameters.