OxfordXML: From Thermograms to Concepts: Using Concept Bottleneck Models to Balance Accuracy and Interpretability in Clinical AI Thermography

This talk will explore how Post-hoc Concept Bottleneck Models (PCBMs) can enhance the interpretability and accuracy of Al in clinical thermography. It introduces concept-based models (CBMs), explains their limitations in real-world use, and shows how PCBMs address these through methods that reveal decision pathways and biases.

The talk will focus on applying PCBMs to the complex field of clinical thermography, emphasizing their value in identifying which guideline-based concepts truly influence risk assessments. Finally, it will highlight how PCBMs support outcome validation and foster deeper clinician engagement through transparent, concept-driven Al.