PromptRad: Knowledge-Enhanced Multi-Label Prompt-Tuning for Low-Resource Radiology Report Labeling

Researchers introduced PromptRad, a knowledge-enhanced prompt-tuning method for radiology report labeling.

It reframes multi-label classification as masked language modeling and enriches labels with UMLS-derived synonyms.

In liver CT report experiments, it beat rule-based and fine-tuning baselines using just 32 labeled examples.

The approach matched GPT-4-like performance more efficiently, making clinical text labeling more data-efficient.