Evaluation of large language models in generating pulmonary nodule follow-up recommendations

Rationale and objectives

To evaluate the performance of large language models (LLMs) in generating clinically follow-up recommendations for pulmonary nodules by leveraging radiological report findings and management guidelines.

Materials and methods

This retrospective study included CT follow-up reports of pulmonary nodules documented by senior radiologists from September 1st, 2023, to April 30th, 2024. Sixty reports were collected for prompting engineering additionally, based on few-shot learning and the Chain of Thought methodology. Radiological findings of pulmonary nodules, along with finally prompt, were input into GPT-4o-mini or ERNIE-4.0-Turbo-8K to generate follow-up recommendations. The AI-generated recommendations were evaluated against radiologist-defined guideline-based standards through binary classification, assessing nodule risk classifications, follow-up intervals, and harmfulness. Performance metrics included sensitivity, specificity, positive/negative predictive values, and F1 score.

Results

On 1009 reports from 996 patients (median age, 50.0 years, IQR, 39.0–60.0 years; 511 male patients), ERNIE-4.0-Turbo-8K and GPT-4o-mini demonstrated comparable performance in both accuracy of follow-up recommendations (94.6 % vs 92.8 %, P = 0.07) and harmfulness rates (2.9 % vs 3.5 %, P = 0.48). In nodules classification, ERNIE-4.0-Turbo-8K and GPT-4o-mini performed similarly with accuracy rates of 99.8 % vs 99.9 % sensitivity of 96.9 % vs 100.0 %, specificity of 99.9 % vs 99.9 %, positive predictive value of 96.9 % vs 96.9 %, negative predictive value of 100.0 % vs 99.9 %, f1-score of 96.9 % vs 98.4 %, respectively.

Conclusion

LLMs show promise in providing guideline-based follow-up recommendations for pulmonary nodules, but require rigorous validation and supervision to mitigate potential clinical risks. This study offers insights into their potential role in automated radiological decision support.