🧠 Site Performance Prediction

This project predicts the likelihood of clinical trial sites completing their studies successfully using historical data from ClinicalTrials.gov. It was developed as a hands-on, end-to-end ML project focused on real-world challenges in clinical trial operations.

---

📦 Project Structure

site-performance-prediction/
├── data/
│   ├── raw/
│   │   └── aact-data/                 # Extracted AACT pipe-delimited files
│   └── site_performance.csv          # Cleaned and enriched dataset
├── notebooks/
│   ├── 01_prepare_site_data.ipynb        # ETL: loads, joins, cleans AACT data
│   └── 02_model_site_performance.ipynb  # Modeling and evaluation
├── README.md                         # Project overview and instructions

---

📊 Problem Statement

Clinical trial site performance (completion vs. dropout) is critical to successful studies. By modeling historical data, we aim to:

• Identify high-risk sites before activation
• Surface factors that influence site success
• Support predictive trial planning

---

🧪 Data Source

Data is sourced from the AACT (Aggregate Analysis of ClinicalTrials.gov) public dataset:

• studies.txt, facilities.txt, sponsors.txt, etc.
• Pipe-delimited format
• Updated monthly by CTTI

---

🚀 How to Run

1. Clone this repo and download the AACT dataset
   (see instructions at AACT Download)
2. Extract AACT files into data/raw/aact-data/
3. Open Jupyter and run:

jupyter notebook notebooks/01_prepare_site_data.ipynb

4. Then run the modeling notebook:

jupyter notebook notebooks/02_model_site_performance.ipynb

---

🤖 Models Used

• Random Forest Classifier
• Logistic Regression
• XGBoost (optional)
• Evaluation via Accuracy, Precision, Recall, F1, ROC-AUC

---

📈 Sample Features

Feature Name	Description
phase	Phase of the clinical trial
enrollment	Target enrollment number
study_type	Interventional, Observational, etc.
location_country	Site location
duration_days	Trial duration
num_sites	Number of sites per study
site_status	Outcome label (Completed/Terminated)

---

📌 Output

• site_performance.csv: final dataset for modeling
• .pkl model file: saved trained model for reuse
• Performance plots and feature importance charts

---

📍 Next Steps

• ✅ Develop an interactive dashboard (e.g., Streamlit or Dash)
• 🔜 Publish a blog post or case study on this work
• 🔜 Add Docker + API for full deployment

---

🛠️ Possible Enhancements

• Incorporate SHAP or LIME for explainability
• Add hyperparameter tuning via Optuna or GridSearchCV
• Include a small synthetic dataset to make the repo self-contained

---

👨‍💻 Author

Andy Waters
Sr. Director of Software Engineering

---

📜 License

MIT (or update as appropriate)