An interpretable and transparent machine learning framework for appendicitis detection in pediatric patients

Area of Science:

• Medical Informatics
• Pediatric Surgery
• Artificial Intelligence in Medicine

Background:

• Appendicitis is a common pediatric condition requiring prompt diagnosis to prevent severe complications like peritonitis and sepsis.
• Current diagnostic methods involve laboratory tests and imaging, with a need for improved accuracy and speed.
• Artificial Intelligence (AI) and machine learning offer promising tools for enhancing medical diagnostics.

Purpose of the Study:

• To develop and evaluate supervised learning models for accurate appendicitis diagnosis in pediatric patients.
• To compare the performance of various hyperparameter tuning techniques for optimizing AI diagnostic models.
• To identify key clinical and imaging markers predictive of appendicitis using explainable AI.

Main Methods:

• Utilized six heterogeneous search techniques (Bayesian Optimization, Hybrid Bat Algorithm, etc.) for hyperparameter tuning.
• Trained and evaluated nine classification models for appendicitis detection in children.
• Applied five explainable AI techniques to interpret model predictions and identify significant diagnostic features.

Main Results:

• The Hybrid Bat Algorithm achieved the highest accuracy (94%) with the customized APPSTACK model.
• Explainable AI identified critical diagnostic markers including length of stay, ultrasonographic vermiform appendix detection, white blood cell count, and appendix diameter.
• The developed AI system demonstrates potential for early and validated appendicitis diagnosis.

Conclusions:

• AI, particularly the Hybrid Bat Algorithm, shows high efficacy in diagnosing pediatric appendicitis.
• Key features like appendix diameter and WBC count are crucial for AI-driven appendicitis detection.
• The proposed AI system can aid clinicians in rapid and reliable diagnosis, complementing existing methods.