DxEvolve

This repository contains the implementation code for DxEvolve.

📂 Directory Structure

To ensure correct data loading, please place the codebase and the dataset directories in the same parent folder:

.
├── DxEvolve                           # Source code directory
└── MIMIC-Clinical-Decision-Making-Dataset  # Dataset directory

Important: The code uses relative paths to access clinical data. Maintaining this parallel structure is mandatory.

📊 Dataset Access

The experiments are conducted using the MIMIC-Clinical-Decision-Making-Dataset.

• Access & Authorization: Please refer to the MIMIC-Clinical-Decision-Making-Dataset GitHub repository for detailed instructions on obtaining authorization and downloading the data.

🚀 Quick Start

Prerequisites

• Python: 3.11
• Environment Manager: Conda (Recommended)

Installation

1. Create the environment:

conda create -n dxevolve python=3.11
conda activate dxevolve

2. Install requirements:

pip install -r requirements.txt

---

🏃 Execution

Run the main process using run.py.

python run.py \
  --model "$MODEL_NAME" \
  --test_num 400

Argument Details:

• --model: The identifier for the model. Available options are:

  • DeepSeek-V3.2
  • GLM-4.7
  • Qwen3-30B-A3B-Instruct-2507
  • Qwen3-235B-A22B-Instruct-2507

• --test_num:

  • 400: Runs evaluation on a standard sample of 400 cases.
  • -1: Doctor Labeled Subset.

---

📝 Notes

• Logs: Results, including agent trajectories and intermediate steps, are saved as .pkl files in the new_logs/ directory.
• Evaluation: Use the evaluation scripts to parse these .pkl files for metrics.