Integrating Large Language Models with Deep Learning for Breast Cancer Treatment Decision Support.

Area of Science:

• Oncology
• Medical Informatics
• Artificial Intelligence

Background:

• Breast cancer's molecular heterogeneity complicates treatment decisions.
• Pathology reports and electronic medical records (EMR) are crucial for treatment planning.
• Standardized and reliable decision support is needed for complex breast cancer cases.

Purpose of the Study:

• To develop an integrated clinical decision support system (CDSS).
• To combine large language model (LLM)-based pathology analysis with deep learning for treatment prediction.
• To support standardized and reliable breast cancer treatment decision-making.

Main Methods:

• Analyzed real-world data from 5015 breast cancer patients.
• Utilized Meta-Llama-3-8B-Instruct for automated extraction of TNM stage and tumor size from pathology reports.
• Integrated extracted pathology data with EMR variables for multi-label classification of 16 treatment combinations using six machine learning models (Decision Tree, Random Forest, GBM, XGBoost, DNN, Transformer).

Main Results:

• Gradient Boosting Machine (GBM) and XGBoost models achieved the highest predictive performance (macro-F1 ≈ 0.88-0.89; AUC = 0.867-0.868) using combined LLM-extracted pathology and EMR features.
• Decision Tree and Random Forest showed moderate, reliable performance (macro-F1 = 0.84-0.86; AUC = 0.849-0.821).
• Deep Neural Network (DNN) and Transformer models yielded lower scores, indicating gradient-boosting ensembles are better suited for tabular medical data.

Conclusions:

• The developed AI-based CDSS enhances accuracy and consistency in breast cancer treatment decision support.
• Integration of automated pathology interpretation and deep learning shows potential for real-world cancer care.
• Gradient-boosting ensemble approaches provide clinically reliable treatment recommendations for breast cancer.