RCT-Reviewer

<p align="center"> <img src="banner_transparent.png" alt="Banner" width="80%"/> </p>

This repository contains the machine learning model weights and artifacts required to run the RCT-Reviewer application. It includes models for Risk of Bias assessment, PICO extraction, and RCT classification.

Acknowledgment

This repository contains joblib-converted model artifacts originally developed in RobotReviewer. The models are redistributed here for ease of deployment in the RCT-Reviewer Streamlit application.

Related

• RCT-Reviewer: https://github.com/aurumz-rgb/RCT-Reviewer
• RCT-Reviewer web: https://rct-reviewer.github.io
• RCT-Reviewer Zenodo: https://zenodo.org/records/20618339
• Original RobotReviewer: https://github.com/ijmarshall/robotreviewer
• RobotReviewer Zenodo: https://zenodo.org/records/6855718

---

📦 Model Formats

This repository strictly provides models in modern, Python-native formats optimized for the RCT-Reviewer application:

• .joblib: Compressed serialized models (typically classifiers or vectorizers).
• .npz: NumPy/Numpy sparse matrices (typically model weights or embeddings).

Note: Legacy .pickle, .pck, and TensorFlow/CNN files are not included in this repository. You may find those here: https://huggingface.co/Aurumz/RCT-Reviewer-pickle

---

🛠️ How to Use

You can load these models directly in Python using joblib and scipy/numpy.

Prerequisites

Ensure you have the necessary libraries installed:

pip install joblib scikit-learn scipy numpy huggingface_hub

1. Downloading Models (Automated)

The RCT-Reviewer application uses the huggingface_hub library to download the models to a local cache directory. You can replicate this behavior using the following snippet:

from huggingface_hub import snapshot_download
from pathlib import Path

# Define cache directory
models_dir = Path.home() / ".cache" / "rct_reviewer" / "models"

# Download the entire repository
snapshot_download(
    repo_id="Aurumz/RCT-Reviewer",
    repo_type="model",
    local_dir=models_dir,
    max_workers=1
)

print(f"Models downloaded to: {models_dir}")

2. Loading .joblib Models

Use the joblib library to load classifier artifacts directly.

import joblib
from huggingface_hub import hf_hub_download

# Example: Downloading a specific joblib model
model_path = hf_hub_download(
    repo_id="Aurumz/RCT-Reviewer", 
    filename="data/bias/bias_classifier.joblib"
)

# Load the model
model = joblib.load(model_path)
print(f"Model loaded successfully: {type(model)}")

3. Loading .npz Weight Files

The .npz files typically contain sparse matrices used for Linear SVM weights or TF-IDF vectors.

import numpy as np
from scipy.sparse import load_npz
from huggingface_hub import hf_hub_download

# Example: Downloading sparse weights
weights_path = hf_hub_download(
    repo_id="Aurumz/RCT-Reviewer", 
    filename="data/rct/rct_svm_weights.npz"
)

# Load the sparse matrix
weights = load_npz(weights_path)

# If it is a standard dense numpy array, use:
# weights = np.load(weights_path)

print(f"Weights shape: {weights.shape}")

📂 File Structure

The artifacts are organized by task within the data directory:

data/
├── bias/           # Risk of Bias models (.npz, .joblib)
├── pico/           # PICO extraction models (.npz)
├── rct/            # RCT classification weights (.npz)
└── vocab/          # Vocabulary and embedding files (.npz)