Automated Pathologic TN Classification Prediction and Rationale Generation From Lung Cancer Surgical Pathology Reports Using a Large Language Model Fine-Tuned With Chain-of-Thought: Algorithm Development and Validation Study.

Background: Traditional rule-based natural language processing approaches in electronic health record systems are effective but are often time-consuming and prone to errors when handling unstructured data. This is primarily due to the substantial manual effort required to parse and extract information from diverse types of documentation. Recent advancements in large language model (LLM) technology have made it possible to automatically interpret medical context and support pathologic staging. However, existing LLMs encounter challenges in rapidly adapting to specialized guideline updates. In this study, we fine-tuned an LLM specifically for lung cancer pathologic staging, enabling it to incorporate the latest guidelines for pathologic TN classification.

Objective: This study aims to evaluate the performance of fine-tuned generative language models in automatically inferring pathologic TN classifications and extracting their rationale from lung cancer surgical pathology reports. By addressing the inefficiencies and extensive parsing efforts associated with rule-based methods, this approach seeks to enable rapid and accurate reclassification aligned with the latest cancer staging guidelines.

Methods: We conducted a comparative performance evaluation of 6 open-source LLMs for automated TN classification and rationale generation, using 3216 deidentified lung cancer surgical pathology reports based on the American Joint Committee on Cancer (AJCC) Cancer Staging Manual8th edition, collected from a tertiary hospital. The dataset was preprocessed by segmenting each report according to lesion location and morphological diagnosis. Performance was assessed using exact match ratio (EMR) and semantic match ratio (SMR) as evaluation metrics, which measure classification accuracy and the contextual alignment of the generated rationales, respectively.

Results: Among the 6 models, the Orca2_13b model achieved the highest performance with an EMR of 0.934 and an SMR of 0.864. The Orca2_7b model also demonstrated strong performance, recording an EMR of 0.914 and an SMR of 0.854. In contrast, the Llama2_7b model achieved an EMR of 0.864 and an SMR of 0.771, while the Llama2_13b model showed an EMR of 0.762 and an SMR of 0.690. The Mistral_7b and Llama3_8b models, on the other hand, showed lower performance, with EMRs of 0.572 and 0.489, and SMRs of 0.377 and 0.456, respectively. Overall, the Orca2 models consistently outperformed the others in both TN stage classification and rationale generation.

Conclusions: The generative language model approach presented in this study has the potential to enhance and automate TN classification in complex cancer staging, supporting both clinical practice and oncology data curation. With additional fine-tuning based on cancer-specific guidelines, this approach can be effectively adapted to other cancer types.