Robust and Interpretable AI for Acute Appendicitis: A Simulation-to-Clinical Validation Pipeline

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Diagnostic Accuracy

Background:

• Accurate diagnosis of acute appendicitis is challenging due to overlapping symptoms and varied presentations.
• Existing diagnostic methods require improvement for reliability and generalizability.
• The need for robust and interpretable AI tools in emergency medicine is critical.

Purpose of the Study:

• To introduce and evaluate the Robust and Interpretable Simulation-Analysis (RISA) framework for assessing diagnostic classifiers.
• To benchmark the performance, robustness, and generalizability of various machine learning algorithms.
• To ensure the interpretability and clinical validity of AI models for acute appendicitis diagnosis.

Main Methods:

• Six supervised learning algorithms were evaluated across 24 simulation scenarios with varying data characteristics.
• Clinical validation was performed on two independent datasets: adult Appendicitis (n=106) and Regensburg Pediatric Appendicitis (n=782).
• Interpretability was assessed using Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP).

Main Results:

• Linear Discriminant Analysis and Support Vector Machine showed balanced performance in simulations; Random Forest excelled in nonlinear scenarios.
• Linear Discriminant Analysis achieved the highest AUC (0.908) on the appendicitis dataset.
• Support Vector Machine demonstrated the highest discrimination (AUC=0.762) on the RPA dataset, with LIME/SHAP confirming clinically relevant biomarkers.

Conclusions:

• The RISA framework provides a reproducible method for developing reliable diagnostic AI.
• It bridges algorithmic robustness with clinical reliability for acute appendicitis diagnosis.
• The study facilitates the implementation of explainable AI in clinical practice.