Using Causal Analysis for Conceptual Deep Learning Explanation.

Area of Science:

• Artificial Intelligence in Medicine
• Machine Learning Explainability
• Clinical Decision Support

Background:

• Trustworthy machine learning (ML) in healthcare requires explainability.
• Explanations should mirror domain expert reasoning and use clinician-friendly terminology.
• Current ML models often lack transparency in clinical applications.

Purpose of the Study:

• To develop an explainable ML approach for healthcare applications.
• To associate hidden ML model units with clinically relevant concepts from radiology reports.
• To provide clinicians with understandable decision rules derived from ML models.

Main Methods:

• Utilized radiology reports from chest X-ray images to define clinical concepts.
• Employed linear sparse logistic regression to discover associations between concepts and hidden model units.
• Applied causal inference and mediation analysis to validate unit influence on classifier outcomes.
• Constructed a low-depth decision tree for translating concepts into clinical decision rules.

Main Results:

• Successfully associated hidden units of a classifier with clinically relevant concepts.
• Identified sparse associations between concepts and hidden units.
• Validated the causal influence of identified units on model predictions.
• Generated a global explanation consistent with clinical knowledge on a large chest X-ray dataset.

Conclusions:

• The proposed method enhances trust in ML models within healthcare.
• The approach translates complex ML models into clinically meaningful explanations for radiologists.
• This work bridges the gap between ML predictions and clinical interpretation in medical imaging.