JMIR Publications

ABSTRACT

Background:

Accurate staging of esophageal cancer is crucial for determining prognosis and guiding treatment strategies, but manual interpretation of radiology reports by clinicians is prone to variability and limited accuracy, resulting in reduced staging accuracy. Recent advances in large language models (LLMs) have shown promise in medical applications, but their utility in esophageal cancer staging remains underexplored.

Objective:

To compare the performance of three publicly available LLMs (INF, Qwen2.5-72B, and LLaMA3.1-70B) and clinicians in automated esophageal cancer staging using free-text radiology reports.

Methods:

This retrospective study included 161 patients from Shanghai Chest Hospital who underwent esophageal cancer surgery from May to September 2024. The dataset consisted of 916 Chinese free-text radiology reports. The reference standard was derived from postoperative pathological staging. Three LLMs determined T classification (T1-T4), N classification (N0-N3), and overall staging (I-IV) using three prompting strategies (Zero-shot, Chain-of-thought, and a proposed Interpretable Reasoning method). The McNemar test and Pearson chi-square test were used for comparisons.

Results:

INF combined with Interpretable Reasoning demonstrated superior performance, achieving an accuracy of 58% in overall staging, which significantly outperformed clinicians (37%, P < .001). For T classification, it achieved an F1 score of 0.67, markedly surpassing clinicians (0.43, P < .001). For N classification, it achieved an F1 score of 0.54, significantly exceeding the performance of clinicians (0.42, P < .001). For overall staging, it attained an F1 score of 0.60, significantly outperforming clinicians (0.39, P < .001).

Conclusions:

The findings demonstrate the potential of LLMs to improve the accuracy and reliability of esophageal cancer staging based on radiology reports. LLMs can serve as valuable tools to support clinicians in complex diagnostic tasks when properly adapted.