Sleep Disorder Risk Prediction

Predict sleep disorder risk using lifestyle and health features from NHANES survey data (2005-2016).

The model can be found and tested via HuggingFace at: https://agalexiul-sleep-risk-model.hf.space/

Table of Contents

• Overview
• Features
• Installation
• Usage
• Project Structure
• Model Details
• Results
• Troubleshooting
• Contributing
• License
• Citation
• Contact

Overview

This project predicts high risk vs low risk for sleep disorders using machine learning on NHANES (National Health and Nutrition Examination Survey) data. It provides:

• Automated data pipeline for NHANES cycles 2005-2016
• Feature engineering from lifestyle and health metrics
• XGBoost classifier with Optuna hyperparameter tuning and MLflow logging
• Interactive Streamlit web interface for risk assessment (see app/streamlit_app.py)

⚠️ Disclaimer: This tool provides risk estimates for research purposes only. It is not a medical diagnostic tool and should not replace clinical evaluation.

Features

• Automated Data Pipeline: Downloads and processes 6 NHANES cycles automatically
• Feature Engineering: 13 features including BMI, exercise, diet, blood pressure, depression score
• Model Performance: ROC-AUC ~0.84-0.94 on test set
• Explainability: SHAP values for feature importance analysis
• Web Interface: User-friendly Streamlit app for predictions
• Reproducible: Version-pinned dependencies and fixed random seeds

Installation

Prerequisites

• Python 3.8 or higher
• pip package manager
• ~1GB free disk space for data
• Internet connection for downloading NHANES data

Step 1: Clone Repository

git clone https://github.com/AlexPi44/sleep-risk-model.git
cd sleep-risk-model

Step 2: Create Virtual Environment

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

Step 3: Install Dependencies

pip install -r requirements.txt

Critical Dependency: pyreadstat>=1.1.5 is required to read NHANES XPT files.

Verify installation:

python -c "import pyreadstat; print('✓ pyreadstat installed successfully')"

Usage

Recommended Pipeline

1. Prepare Data

Run the full data preparation pipeline (downloads, merges, feature engineering):

python notebooks/01_data_prep.py

This will:

• Download NHANES data (~10-30 minutes)
• Merge and engineer features (~1-2 minutes)
• Output: data/processed/merged_clean.csv

2. Tune and Train Model

Run Optuna hyperparameter tuning and train the best XGBoost model:

python src/tune_xgb_optuna.py --n-trials 40 --tracking-uri http://localhost:5000 --experiment sleep_risk_optuna

This will:

• Tune and train the model
• Save model artifacts to models/:
  • xgb_best_model.pkl (final model)
  • feature_names.pkl (feature order)
  • xgb_cat_maps.json (categorical mappings)

3. Launch the Web App

streamlit run app/streamlit_app.py

Open your browser to http://localhost:8501.

---

Manual Steps (Advanced)

You can also run the following scripts directly for more control:

• Download NHANES data:
  python src/check_nhanes_downloads.py --out data/raw --report reports --convert
• Merge and engineer features:
  python src/merge_and_engineer.py --input data/raw --output data/processed/merged_clean.csv

Project Structure

sleep-risk-model/
├── app/
│   └── streamlit_app.py           # Streamlit web interface
├── src/
│   ├── check_nhanes_downloads.py  # Data downloader
│   ├── merge_and_engineer.py      # Feature engineering
│   └── tune_xgb_optuna.py         # Optuna tuning & training
├── notebooks/
│   ├── 01_data_prep.ipynb         # Jupyter notebook
│   └── 01_data_prep.py            # Python script version (full pipeline)
├── data/
│   ├── raw/                       # Downloaded NHANES files
│   └── processed/                 # Merged dataset
├── models/                        # Trained model artifacts
├── reports/                       # Download reports
├── requirements.txt               # Python dependencies
├── LICENSE                        # MIT License
└── README.md                      # This file

Model Details

Target Variable

Participants are labeled high risk (1) if any of:

• Doctor-diagnosed sleep disorder (SLQ060 == 1)
• Sleep duration < 7 or > 9 hours (SLD012)
• Excessive daytime sleepiness ≥ 16 times/month (SLQ120)

Otherwise labeled low risk (0) if sleep duration is 7-9 hours with no diagnosis.

Features (13 total)

Non-modifiable:

• Age, Sex

Modifiable lifestyle factors:

• BMI (Body Mass Index)
• Exercise (minutes/week)
• Diet: calories, fiber, added sugar, caffeine (daily averages)
• Alcohol consumption (drinks/week)
• Current smoker (yes/no)
• Depression score (PHQ-9: 0-27)
• Blood pressure (systolic, diastolic)

Model Architecture

• Algorithm: XGBoost Classifier
• Parameters: 100 trees, max depth 6, learning rate 0.1
• Preprocessing: StandardScaler normalization
• Train/Test Split: 80/20, stratified by target
• Evaluation: ROC-AUC, PR-AUC, confusion matrix

Feature Engineering

• Exercise: Moderate minutes + 2× vigorous minutes per week
• Diet: Average of two 24-hour dietary recalls
• Depression: Sum of 9 PHQ-9 questionnaire items
• Alcohol: Daily drinks × 7 = weekly consumption

Results

Model Performance

Metric	Value
ROC-AUC	0.84-0.94
PR-AUC	0.79-0.84
Training Samples	~15,000-20,000
Test Samples	~3,000-5,000

Top Features (by SHAP importance)

1. Age
2. Depression score
3. BMI
4. Sleep duration (embedded in target)
5. Blood pressure

Troubleshooting

Issue	Solution
ModuleNotFoundError: No module named 'pyreadstat'	pip install pyreadstat>=1.1.5
FileNotFoundError: XPT file not found	Run downloader: python src/check_nhanes_downloads.py ...
404 errors during download	Check CDC NHANES website: https://www.cdc.gov/nchs/nhanes/
merged_clean.csv not found	Run merge script: python src/merge_and_engineer.py ...
Streamlit shows "No model found"	Verify 2 files exist in models/ directory
Very few training samples (<1000)	Check reports/ for missing sleep variables
Download timeout	Increase REQUEST_TIMEOUT in check_nhanes_downloads.py (line 56)

Data Source

NHANES (National Health and Nutrition Examination Survey)

• Provider: U.S. Centers for Disease Control and Prevention (CDC)
• Cycles: 2005-2006, 2007-2008, 2009-2010, 2011-2012, 2013-2014, 2015-2016
• License: Public domain, free for use
• Documentation: https://www.cdc.gov/nchs/index.html

Limitations

• Not a diagnostic tool - Predicts risk only, not clinical diagnosis
• U.S. population - Trained on NHANES; may not generalize internationally
• Self-reported data - Subject to recall and reporting bias
• Data age - Trained on 2005-2016 data; modern populations may differ
• Missing data - ~70% of samples dropped due to incomplete sleep questionnaires
• Class imbalance - Consider using SMOTE or class weights for production use

Contributing

Contributions are welcome! Please:

1. Fork the repository
2. Create a feature branch (git checkout -b feature/improvement)
3. Commit changes (git commit -am 'Add new feature')
4. Push to branch (git push origin feature/improvement)
5. Open a Pull Request

Development Setup

pip install -r requirements.txt
pip install pytest black flake8  # Development dependencies

License

This project is licensed under the MIT License - see the LICENSE file for details.

NHANES data is public domain courtesy of the CDC.

Citation

If you use this code in your research, please cite:

@software{sleep_risk_model_2025,
  author = {{Alexandru Paicu}},
  title = {Sleep Disorder Risk Prediction using NHANES Data},
  year = {2025},
  url = {https://github.com/AlexPi44/sleep-risk-model},
  version = {1.0.0}
}

Contact

Project Maintainer: Alexandru Paicu Email: paicualexandru44@gmail.com
GitHub: https://github.com/AlexPi44/sleep-risk-model

---

Acknowledgments

• CDC NHANES team for public data
• scikit-learn, XGBoost, and SHAP contributors
• Streamlit for the web framework