This is repo of MICCAI 2025 paper 'Prolog-Driven Rule-Based Diagnostics with Large Language Models for Precise Clinical Decision Support'

├── data/ # Example of Anonymized Dataset
├── figures/ # Figures
├── Prompts.py # Prompts used in Stage 1 and Stage 2
├── Stage1_extracItnfo.py # Workflow of Stage 1
├── Stage2_prologGenerate.py # Workflow of Stage 2
├── Stage2_verifyCode.py # Workflow of Stage 2
├── prologInference.py # Inference Code of Prolog
└── README.md 

• Framework: The overall framework is illustrated as follows:

• LLM-Powered Information Extraction: Leverages LLMs to understand and extract key rules and patient facts from text.
• Automated Prolog Generation: Translates extracted information into executable Prolog code.
• LLM-Based Code Verification: Employs an LLM to review and refine the generated Prolog code, improving its reliability.
• Multi-Stage Pipeline: A clear, step-by-step workflow from raw data to final inference.

• Python Libraries:

  • vllm: For high-throughput LLM inference. Requires a compatible GPU (e.g., NVIDIA) and CUDA toolkit.
  • jsonlines: For handling JSONL files.
  • pyswip: Python-Prolog Interface

  pip install vllm jsonlines pyswip

• External Software:

  • SWI-Prolog: The prologInference.py script requires a SWI-Prolog interpreter to be installed and available in the system's PATH.

First, ensure all dependencies are installed. Then, run the scripts in sequence.

# --- Configuration ---
# Path to the local LLM model compatible with vLLM
MODEL_PATH="/path/to/your/Llama-3-8B-Chat" 
# Path to the input data
INPUT_DATA="data/patient.jsonl"

# --- Output file paths ---
EXTRACT_OUTPUT="data/data_extract.jsonl"
PROLOG_OUTPUT="data/prolog.jsonl"
PROLOG_VERIFIED_OUTPUT="data/prolog_verified.jsonl"
FINAL_RESULTS="data/results.jsonl"

# Stage 1: Extract Rules and Facts
python submit/Stage1_extractInfo.py \
    --model_path $MODEL_PATH \
    --input_path $INPUT_DATA \
    --output_path $EXTRACT_OUTPUT

# Stage 2a: Generate Prolog Code
python submit/Stage2_prologGenerate.py \
    --model_path $MODEL_PATH \
    --input $EXTRACT_OUTPUT \
    --output $PROLOG_OUTPUT

# Stage 2b: Run Prolog Inference
python submit/prologInference.py \
    --input $PROLOG_OUTPUT \
    --output $PROLOG_RESULTS

# Stage 2c: Verify Prolog Code
python submit/Stage2_verifyCode.py \
    --model_path $MODEL_PATH \
    --input $PROLOG_OUTPUT \
    --output $PROLOG_VERIFIED_OUTPUT

# Stage 2d: Run Prolog Inference
python submit/prologInference.py \
    --input $PROLOG_VERIFIED_OUTPUT \
    --output $FINAL_RESULTS

We extend our sincere gratitude to the following open-source projects and communities for their invaluable contributions:

• SWI-Prolog – The powerful Prolog engine that enabled efficient symbolic reasoning in our work.
• PySwip – The Python-SWI-Prolog bridge that made seamless integration possible.
• vLLM – The high-performance LLM serving framework that accelerated our inference pipelines.

We deeply appreciate the efforts of their developers and maintainers.

@inproceedings{tan2026prolog,
  title={Prolog-Driven Rule-Based Diagnostics with Large Language Models for Precise Clinical Decision Support},
  author={Tan, Xiaoyu and Li, Bin and Xu, Weidi and Qu, Chao and Chu, Wei and Xu, Yinghui and Qi, Yuan and Qiu, Xihe},
  booktitle={Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
  pages={413--423},
  year={2026},
  publisher={Springer Nature Switzerland},
  address={Cham},
  isbn={978-3-032-05127-1}
}