New AI Model Improves Preoperative Severity Prediction for Pediatric Appendicitis: Study

"The model successfully classified appendicitis into five severity grades, achieving an accuracy of 70.1% and an AUROC of 0.77. It also effectively distinguished between non-perforated and perforated appendicitis, with a negative predictive value (NPV) of 82.8% and a positive predictive value (PPV) of 56.4%," the researchers wrote in the Journal of Pediatric Surgery.

Aylin Erman, Department of Computer Science, McGill University, Montreal, QC, and colleagues aimed to assess the effectiveness of machine learning (ML) algorithms in enhancing the preoperative diagnosis of acute appendicitis in children, with an emphasis on accurately predicting disease severity.

For this purpose, the researchers retrieved an anonymized clinical and operative dataset from the medical records of children who underwent emergency appendectomy between 2014 and 2021. They developed a machine learning (ML) pipeline to preprocess the dataset and create algorithms that predict five appendicitis severity grades: 1 - non-perforated, 2 - localized perforation, 3 - abscess, 4 - generalized peritonitis, and 5 - generalized peritonitis with abscess.

Missing values were imputed, and infrequent classes were addressed using upsampling techniques. Various classifier models were tested, with the best combination of imputation strategy, class balancing, and classification model selected based on validation performance. The model’s explainability was confirmed by a pediatric surgeon and compared with another pediatric appendicitis severity prediction tool.

The study led to the following findings:

• The study analyzed a retrospective cohort of 1,980 patients (60.6% male, average age 10.7 years).

• The distribution of appendicitis severity grades in the cohort was: grade 1 (70%), grade 2 (8%), grade 3 (7%), grade 4 (7%), and grade 5 (8%).

• The researchers tested every combination of 6 imputation strategies, seven class-balancing techniques, and five classification models.

• The best-performing machine learning pipeline achieved:

• 82.8% NPV and 56.4% PPV in distinguishing non-perforated from perforated appendicitis.

• 70.1% accuracy and an AUROC of 0.77 in differentiating between severity grades.

• A comparison with another pediatric appendicitis severity prediction tool showed:

• 71.4% accuracy, AUROC of 0.54, and NPV/PPV of 71.8%/64.7%.

In conclusion, the study demonstrates that the machine learning model for predicting appendiceal perforation outperforms models predicting the full spectrum of appendicitis grades. The key variables identified by the model align with clinical experience and existing literature, indicating that the ML approach successfully identified relevant patterns in the data. Notably, this model surpasses other pediatric appendicitis prediction tools in performance.

"Our ML model for grade prediction represents a novel method for assessing appendicitis severity in children preoperatively. Following further validation and clinical testing, it has the potential to enable personalized treatment and optimize resource allocation in managing pediatric appendicitis," the researchers concluded.

Reference:

Erman, A., Ferreira, J., Ashour, W. A., Guadagno, E., St-Louis, E., Emil, S., Cheung, J., & Poenaru, D. (2025). Machine-learning-assisted preoperative prediction of pediatric appendicitis severity. Journal of Pediatric Surgery, 162151. https://doi.org/10.1016/j.jpedsurg.2024.162151