Machine learning for chemotherapy decision-making in breast cancer using large language model

Area of Science:

• Oncology
• Biostatistics
• Artificial Intelligence

Background:

• Breast cancer chemotherapy decisions are complex due to patient heterogeneity and clinical practice variations.
• Accurate treatment recommendations are crucial for optimizing patient outcomes.

Purpose of the Study:

• To develop a hybrid framework integrating machine learning (ML), causal reasoning, and large language models (LLMs) for improved breast cancer chemotherapy recommendations.
• To enhance personalized and transparent treatment decision support in oncology.

Main Methods:

• Utilized the METABRIC dataset with eleven pre-treatment clinicopathologic variables.
• Developed a Random Forest classifier and compared it with baseline ML models.
• Estimated individualized treatment benefit using causal survival analysis and employed GPT-4 for clinical rationale generation.

Main Results:

• The Random Forest model achieved an AUC of 0.91, surpassing benchmark models.
• Causal analysis revealed heterogeneous treatment benefits and identified patient groups for potential chemotherapy de-escalation.
• GPT-4 demonstrated moderate agreement with the Random Forest and highlighted clinically relevant factors, enhancing interpretability.

Conclusions:

• The hybrid framework combining predictive ML, causal survival modeling, and LLM-based rationale generation offers a novel approach for personalized chemotherapy decision support.
• This integrated strategy promises to improve transparency and effectiveness in oncological treatment planning.