O2170: Machine-learning-assisted preoperative prediction of pediatric appendicitis severity

Background: This study evaluates the effectiveness of Machine Learning (ML) algorithms in improving the preoperative diagnosis of acute appendicitis in children, focusing on the accurate prediction of the presence and severity of complications.

Methods: We collected a preoperative clinical and intra-operative dataset from children undergoing emergency appendectomy between 2014 and 2021 at one tertiary institution. We developed an ML pipeline that pre-processed the dataset and developed algorithms to predict the presence and severity of complications, using 5 appendicitis grades (1 - non-perforated, 2 - localized perforation, 3 - abscess, 4 - generalized peritonitis, and 5 - generalized peritonitis with abscess).
Several imputation strategies were used for substituting missing data and upsampling techniques for sample imbalance (due to infrequent appendicitis grades, or classes). To predict the presence and severity of complications, 2 different computational approaches were tested. The direct approach generated one classifier algorithm directly predicting the appendicitis grade. The indirect approach developed 3 classifiers to predict 1. postoperative intra-abdominal abscess (none, single, multiple), 2. peritonitis (none, localized, generalized) and 3. perforation (present, not present) — which were then combined to deterministically identify the appendicitis grade. Standard ML classifiers including logistic regression, k-nearest neighbors, random forest, decision tree, and ensemble models were tested for each approach. The best combination of imputation strategy, class balancing technique, prediction approach, and classifier were then identified based on the performance of each model in the final validation stage. The explainability of each derived model was verified by a pediatric surgeon.

Results: The study included 1576 patients (60.8% males, average age 14 years). The appendicitis grades in the cohort were as follows: grade 1 - 72%; grade 2 - 4%; grade 3 - 6%; grade 4 - 7%; and grade 5 - 11%. Every combination of the 6 imputation strategies, 7 class balancing techniques, 2 prediction approaches and 5 classification models was tested. The best-performing combined ML pipeline distinguished well simple from perforated appendicitis (90.1% NPV, 59.5% PPV), and differentiated between severity grades with 81.8% accuracy. The indirect approach, predicting first abscess, peritonitis and perforation, gave the best result. Model learning for intra-abdominal abscess was based on duration of symptoms, fever, % neutrophils, nausea, diarrhea and temperature; for peritonitis based on fever, % neutrophils, ultrasound identification of appendix and intra-abdominal abscess, guarding and pain migration; and for perforation based on duration of symptoms, % neutrophils, fever, white blood cell count, appendix diameter, and age (all in decreasing order).

Conclusion: The ML model developed is the first of this type in pediatric appendicitis, offering a novel approach for assessing appendicitis severity preoperatively. Following external validation and silent clinical testing, this ML model has the potential to enable personalized severity-based treatment of pediatric appendicitis and to optimize resource allocation for its management.