Beyond Binary Cutoffs: An Explainable Machine Learning Framework for Individualized Diagnostic Reasoning in Suspected...

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Decision Support Systems
• Medical Diagnostics

Background:

• Emergency department (ED) evaluation of suspected urolithiasis (kidney stones) often uses non-contrast CT, but not all patients require imaging.
• Existing clinical prediction rules use binary indicators, limiting insight into individual risk factors and diagnostic uncertainty.
• There's a need for tools that provide individualized diagnostic reasoning for suspected urolithiasis.

Purpose of the Study:

• To develop and evaluate an explainable machine learning (ML) framework for suspected urolithiasis diagnosis.
• To quantify the diagnostic contribution of sequential testing stages using a Shannon entropy-based framework.
• To provide individualized risk insights by analyzing continuous clinical and laboratory features.

Main Methods:

• Retrospective analysis of 1000 ED patients with suspected urolithiasis.
• Trained a gradient boosting classifier on 17 continuous features, comparing it to binary-thresholded models and a reference score.
• Utilized Shapley values for prediction explanation and Shannon entropy for quantifying information gain at each testing stage.

Main Results:

• The 17-feature ML model achieved higher AUC (0.771) than the reference score (0.723).
• Key predictors included red blood cell count, age, pain duration, and prior stone history; non-linear contributions were observed.
• Dipstick urinalysis provided the largest marginal information gain (6.1%), followed by physical examination (2.3%). C-reactive protein was a significant negative predictor.

Conclusions:

• Explainable ML can reveal which features drive risk, their continuous behavior, and how much uncertainty is resolved at each diagnostic step.
• The Shapley-based explanations and entropy framework offer a structured approach to individualized diagnostic reasoning for urolithiasis.
• This framework serves as an interpretive adjunct to clinical judgment and imaging, not a standalone triage tool.