ScaFE: Data-Efficient Scar Classification with LLM-Generated Clinical Feature Programs

TL;DR - ScaFE uses an LLM to generate deterministic clinical feature programs for local scar-image classification, improving auditability and data governance. It outperforms BiomedCLIP across hospitals while remaining effective with limited labeled data.

• Achieves 81.0% site-macro balanced accuracy on 600 images from three hospitals, 10 points above BiomedCLIP.
• Retains 72.0% accuracy with 10% of development data, maintaining an 11.8-point lead.
• Keeps raw images and patient-level outputs local; only aggregate metrics and SHAP summaries support LLM-guided refinement.
• Iteration increases executable programs from 66.7% to 95.0%, with verified evidence for 91.7% of final features.