A machine learning pipeline differentiated between pediatric appendicitis severity grades with 70.1% accuracy and 0.77 AUROC, compared to 71.4% accuracy and 0.54 AUROC for an existing tool.
Cohort (n=1,980)
Does a machine learning pipeline improve the preoperative prediction of appendicitis severity in children compared to existing tools?
A novel machine learning pipeline accurately predicts pediatric appendicitis severity preoperatively, outperforming existing prediction tools.
Effect estimate: AUROC 0.77 vs 0.54
Absolute Event Rate: 70.1% vs 71.4%
PURPOSE: This study evaluates the effectiveness of machine learning (ML) algorithms for improving the preoperative diagnosis of acute appendicitis in children, focusing on the accurate prediction of the severity of disease. METHODS: An anonymized clinical and operative dataset was retrieved from the medical records of children undergoing emergency appendectomy between 2014 and 2021. We developed an ML pipeline that pre-processed the dataset and developed algorithms to predict 5 appendicitis grades (1 - non-perforated, 2 - localized perforation, 3 - abscess, 4 - generalized peritonitis, and 5 - generalized peritonitis with abscess). Imputation strategies were used for missing values and upsampling techniques for infrequent classes. Standard classifier models were tested. The best combination of imputation strategy, class balancing technique and classification model was chosen based on validation performance. Model explainability was verified by a pediatric surgeon. Our model's performance was compared to another pediatric appendicitis severity prediction tool. RESULTS: The study used a retrospective cohort including 1980 patients (60.6 % males, average age 10.7 years). Grade of appendicitis in the cohort was as follows: grade 1-70 %; grade 2-8 %; grade 3-7 %; grade 4-7 %; grade 5-8 %. Every combination of 6 imputation strategies, 7 class-balancing techniques, and 5 classification models was tested. The best-performing combined ML pipeline distinguished non-perforated from perforated appendicitis with 82.8 ± 0.2 % NPV and 56.4 ± 0.4 % PPV, and differentiated between severity grades with 70.1 ± 0.2 % accuracy and 0.77 ± 0.00 AUROC. The other pediatric appendicitis severity prediction tool gave an accuracy of 71.4 %, AUROC of 0.54 and NPV/PPV of 71.8/64.7. CONCLUSION: Prediction of appendiceal perforation outperforms prediction of the continuum of appendicitis grades. The variables our models primarily rely on to make predictions are consistent with clinical experience and the literature, suggesting that the ML models uncovered useful patterns in the dataset. Our model outperforms the other pediatric appendicitis prediction tools. The ML model developed for grade prediction is the first of this type, offering a novel approach for assessing appendicitis severity in children preoperatively. Following external validation and silent clinical testing, this ML model has the potential to enable personalized severity-based treatment of pediatric appendicitis and optimize resource allocation for its management.
Erman et al. (Mon,) conducted a cohort in Pediatric acute appendicitis (n=1,980). Machine learning pipeline vs. Another pediatric appendicitis severity prediction tool was evaluated on Differentiation between severity grades (accuracy) (AUROC 0.77 vs 0.54). A machine learning pipeline differentiated between pediatric appendicitis severity grades with 70.1% accuracy and 0.77 AUROC, compared to 71.4% accuracy and 0.54 AUROC for an existing tool.