Pre-operative T-stage discrimination in gallbladder cancer using machine learning and DeepSeek-R1

Area of Science:

• Oncology
• Medical Informatics
• Machine Learning

Background:

• Gallbladder cancer (GBC) often presents with non-specific symptoms, delaying diagnosis.
• Early detection of GBC T-stages is crucial for effective treatment planning and patient outcomes.

Purpose of the Study:

• To assess if routine blood biomarkers can differentiate early GBC T-stages using machine learning.
• To compare the T-stage discrimination performance of a large language model (DeepSeek-R1) using radiology reports alone versus reports plus biomarker data.

Main Methods:

• Retrospective analysis of 232 GBC patients, classifying T1 and T2 stages.
• Training machine learning models (Random Forest, SVC, XGBoost, LightGBM) with seven blood biomarkers (NLR, MLR, PLR, CEA, CA19-9, CA125, AFP).
• Evaluating DeepSeek-R1's performance in classifying T1 vs. T2 GBC using radiology reports with and without biomarker data.

Main Results:

• Biomarker-based machine learning models showed poor T-stage discrimination, with performance near random chance.
• Synthetic Minority Over-sampling Technique (SMOTE) did not improve performance on the independent test set.
• DeepSeek-R1 achieved 89.6% accuracy using only radiology report text, identifying features like 'gallbladder wall thickening' for T2 classification; biomarker data did not enhance this accuracy.

Conclusions:

• Routine blood biomarkers are ineffective for early GBC T-stage discrimination.
• A radiology text-driven large language model demonstrates high accuracy and interpretability, offering potential for clinical decision support in GBC.
• Further multi-center studies are needed to validate these findings in larger GBC cohorts.