In-Context Learning with Large Language Models: A Simple and Effective Approach to Improve Radiology Report Labeling.

OBJECTIVES

This study assessed the effectiveness of in-context learning using Generative Pre-trained Transformer-4 (GPT-4) for labeling radiology reports.

METHODS

In this retrospective study, radiology reports were obtained from the Medical Information Mart for Intensive Care III database. Two structured prompts-the "basic prompt" and the "in-context prompt"- were compared. An optimization experiment was conducted to assess consistency and the occurrence of output format errors. The primary labeling experiments were performed on 200 unseen head computed tomography (CT) reports for multilabel classification of predefined labels (Experiment 1) and on 400 unseen abdominal CT reports for multi-label classification of actionable findings (Experiment 2).

RESULTS

The inter-reader accuracies in Experiments 1 and 2 were 0.93 and 0.84, respectively. For multi-label classification of head CT reports (Experiment 1), the in-context prompt led to notable increases in F1-scores for the "foreign body" and "mass" labels (gains of 0.66 and 0.22, respectively). However, improvements for other labels were modest. In multi-label classification of abdominal CT reports (Experiment 2), in-context prompts produced substantial improvements in F1-scores across all labels compared to basic prompts. Providing context equipped the model with domain-specific knowledge and helped align its existing knowledge, thereby improving performance.

CONCLUSIONS

Incontext learning with GPT-4 consistently improved performance in labeling radiology reports. This approach is particularly effective for subjective labeling tasks and allows the model to align its criteria with those of human annotators for objective labeling. This practical strategy offers a simple, adaptable, and researcher-oriented method that can be applied to diverse labeling tasks.