Large Language Model and Knowledge Graph-Driven AJCC Staging of Prostate Cancer Using Pathology Reports

Area of Science:

• Oncology
• Medical Informatics
• Artificial Intelligence

Background:

• Accurate cancer staging is crucial for treatment planning and prognosis.
• Manual review of pathology reports for staging is time-consuming and prone to errors.
• Automating the American Joint Committee on Cancer (AJCC) staging process for radical prostatectomy pathology reports is needed.

Purpose of the Study:

• To develop an automated AJCC staging system for radical prostatectomy pathology reports.
• To leverage large language model (LLM)-based information extraction and knowledge graph validation for staging.
• To enhance clinical decision support in oncology through privacy-protected, automated staging.

Main Methods:

• Utilized pathology reports from 152 radical prostatectomy patients for initial development.
• Employed GPT-4.1 for zero-shot extraction of five key pathological parameters.
• Constructed a knowledge graph for rule-based AJCC staging and consistency validation, with LLM performance evaluated on 16 parameters.
• Validated the system on an external dataset of 88 radical prostatectomy patients from The Cancer Genome Atlas (TCGA).

Main Results:

• Achieved high accuracy (0.973) and F1-score (0.986) for information extraction on the internal dataset, and strong performance on external validation (0.938 accuracy, 0.968 F1-score).
• Demonstrated robust AJCC staging classification with macro-averaged F1-scores of 0.930 (internal) and 0.833 (external).
• Knowledge graph validation identified data inconsistencies in 3.3% of cases.

Conclusions:

• Automated AJCC staging is feasible by integrating LLM information extraction and knowledge graph validation.
• The developed system offers privacy-protected clinical decision support for cancer staging.
• The system shows potential for extensibility to other oncologic domains.