GatorOnco: An agentic generative large language model for treatment planning of colorectal cancer

Overview

Colorectal cancer (CRC) treatment planning requires highly personalized and to remain aligned with rapidly evolving clinical guidelines. However, general-purpose LLMs can struggle with clinical currency and safety in oncology. Here we introduce GatorOnco, an agentic LLM system that generates safety-first, guideline-grounded CRC treatment plans from heterogeneous clinical context. GatorOnco combines system-scale domain adaptation and model merging to preserve instruction-following while strengthening oncology competence, hierarchical reward–based reinforcement learning that prioritizes correct high-level modality decisions before regimen-level details, and agentic retrieval that dynamically incorporates up-to-date NCCN guidance during inference. In a blinded evaluation of 79 real-world CRC cases reviewed by five board-certified oncologists, GatorOnco outperformed open-source baselines and achieved expert-level parity on correctness, guideline adherence, and safety, while improving completeness.

Pipeline

1. Preprocess raw notes.
2. Construct training/evaluation datasets.
3. Build guideline-derived knowledge artifacts.
4. Run model inference and evaluation.

Repository Structure

• preprocess/: note de-duplication, keyword filtering, and temporal filtering.
• dataset_construction/: patient case rewriting, treatment plan rewriting, and reasoning trace generation.
• vector_database_construction/: guideline PDF parsing and metadata extraction for vector database building.
• evaluation/: inference scripts and automatic/human evaluation scripts.
• figs/: figures for system and results.
• flowchart.md: mermaid pipeline diagram.

Figures

System overview

Dataset construction

Treatment label accuracy

Human evaluation

System Requirements

• Python 3.10+ recommended.
• CUDA GPU recommended for vllm and local large-model inference.
• Sufficient RAM/VRAM based on chosen model size.

Installation

git clone <your-github-url>/GatorOnco.git
cd GatorOnco

python -m venv .venv
source .venv/bin/activate
pip install -U pip
pip install -r requirements.txt

Quick Reproduction

# 1) preprocess
bash preprocess/run_all_preprocess.sh

# 2) dataset construction
python dataset_construction/pt_case_rewrite.py --input_csv <csv> --output_csv <csv>
python dataset_construction/plan_rewrite.py --input_csv <csv> --output_csv <csv>
python dataset_construction/reasoning_trace_generation.py --input_csv <csv> --output_csv <csv>

# 3) vector database construction
python vector_database_construction/pdf_to_markdown_jsonl.py \
  --pdf <guideline.pdf> \
  --md_out <guideline.md> \
  --jsonl_out <guideline.jsonl>
python vector_database_construction/extract_metadata_from_jsonl.py

# 4) inference and evaluation
python evaluation/infer_treatment_from_evidence.py --input_csv <csv> --output_jsonl <jsonl>
python evaluation/evaluate_treatment_accuracy.py --gold <csv> --pred <jsonl>
python evaluation/evaluate_treatment_predictions.py <predictions.csv>

Data and Models

• Replace all <csv>, <jsonl>, and model paths with your local files.
• Keep private or licensed clinical data outside this repository.

Citation

If you use this repository, please cite the accompanying paper.

@article{gatoronco2026,
  title={GatorOnco: Agentic Pipeline for Oncology Treatment Plan Generation and Evaluation},
  author={TBD},
  journal={TBD},
  year={2026}
}