To propose a hybrid framework integrating machine learning, causal reasoning, and large language models to improve chemotherapy treatment recommendations for breast cancer.
Approach:
Data Utilization: The METABRIC dataset was used to select eleven pre-treatment clinicopathologic variables.
Model Development: A Random Forest classifier was developed and compared with baseline machine learning models.
Causal Analysis: Individualized treatment benefit was estimated through inverse probability-weighted causal survival analysis.
Rationale Generation: GPT-4 was employed using few-shot prompting to generate clinical rationales.
Key Findings:
GPT-4 showed moderate agreement with the Random Forest classifier (Cohen's κ = 0.13), indicating limited concordance while highlighting clinically relevant factors.
Interpretation:
The proposed framework combines predictive machine learning, causal survival modeling, and LLM-based rationale generation to enhance personalized and transparent chemotherapy decision support in oncology.
Limitations:
Existing studies on LLMs often remain descriptive or concordance-based, lacking direct patient-level outcome assessments related to the proposed framework.
Conclusion:
The hybrid framework offers a promising approach for improving chemotherapy decision-making in breast cancer by integrating predictive capabilities with patient-level reasoning.
by Md Serajun Nabi, Dema Yuden, Thinley Yeshey Choden, S. M. Asiful Islam Saky, Hasanul Bannah, Md Sabbir Hossen, Mohammad Faizal Ahmad Fauzi, Hezerul Bin Abdul Karim
