A QUEST for Model Assessment: Identifying Difficult Subgroups via Epistemic Uncertainty Quantification.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Medical Data Analysis

Background:

• Uncertainty quantification (UQ) in machine learning (ML) offers insights into model reliability and builds trust.
• Well-calibrated UQ links high uncertainty to increased classification errors.
• Current UQ primarily focuses on individual prediction certainty.

Purpose of the Study:

• To investigate if rules explaining ML model uncertainty can identify subgroups with varying performance.
• To determine if UQ can provide a global understanding of model behavior across patient subpopulations.
• To evaluate the utility of UQ beyond individual prediction confidence.

Main Methods:

• Developing rule-based explanations for ML model uncertainty.
• Applying techniques to deep neural networks and gradient boosting ensembles.
• Evaluating methods on benchmark and real-world medical datasets.

Main Results:

• Generated rules successfully delineated subgroups with high and low classification uncertainty.
• These rules corresponded to subgroups where the models exhibited distinct performance levels.
• The approach demonstrated effectiveness across different ML architectures and datasets.

Conclusions:

• Explaining ML uncertainty via rules can globally characterize model performance across patient subgroups.
• UQ's utility extends from individual prediction assessment to understanding model behavior on subpopulations.
• This method enhances trust and interpretability of ML models in medical applications.