Sleep Disorder Classification

A reproducible machine-learning project that classifies a record as Healthy, Insomnia, or Sleep Apnea from sleep, lifestyle, and physiological features.

The deployable artifact is one scikit-learn pipeline containing feature engineering, missing-value handling, encoding, scaling, and classification. Training and prediction therefore apply exactly the same transformations.

Results

Models were compared with stratified, group-aware five-fold cross-validation. Rows with identical predictive features stay in the same fold, preventing duplicate patient profiles from appearing in both training and validation. Selection was based on mean macro F1, which gives equal importance to all three classes.

Model	Macro F1	Mean accuracy	Fold accuracy SD
Logistic Regression	0.863	0.891	0.059
Support Vector Machine	0.884	0.909	0.039
Gradient Boosting	0.860	0.893	0.042

Support Vector Machine achieved the highest mean macro F1 in this comparison and was fitted on the full dataset for the saved pipeline. These are internal cross-validation estimates, not results from an independent external test set.

What Was Corrected

• The model now saves preprocessing and classification as one pipeline.
• Training and prediction use identical feature transformations.
• Person ID is excluded because it is an identifier, not a clinical feature.
• Blood pressure is parsed into systolic and diastolic values, and severe hypertension is checked before the broader stage-2 condition.
• Categorical features use an encoder that safely handles unseen values.
• Numeric and categorical missing values are imputed inside each validation fold, preventing preprocessing leakage.
• Identical feature rows are assigned to the same validation group to prevent duplicate-profile leakage.
• Model selection uses stratified, group-aware five-fold cross-validation instead of one train/test split.

Repository Structure

sleep-disorder-classification/
├── data/
│   └── README.md
├── models/
│   └── sleep_disorder_pipeline.joblib
├── reports/
│   └── model_comparison.png
├── src/
│   ├── train.py
│   ├── predict.py
│   └── preprocessing.py
├── tests/
│   ├── test_prediction.py
│   ├── test_preprocessing.py
│   └── test_training.py
├── .github/workflows/ci.yml
├── .dockerignore
├── .gitignore
├── Dockerfile
├── LICENSE
├── README.md
└── requirements.txt

Dataset

The project uses the Sleep Health and Lifestyle Dataset. The local training file contains 374 records and 13 columns. Missing target values are interpreted as Healthy.

The dataset is published separately under CC0: Public Domain. It is not covered by this repository's MIT License. Download instructions are available in data/README.md; CSV files are intentionally excluded from Git.

Installation

Python 3.11 or newer is recommended.

git clone https://github.com/Dochikhoa2006/Sleeping-Disorder-Analysis.git
cd Sleeping-Disorder-Analysis
python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -r requirements.txt

On Windows PowerShell, activate the environment with:

.venv\Scripts\Activate.ps1

Train

Place the dataset at data/Sleep_health_and_lifestyle_dataset.csv, then run:

python -m src.train

This command compares the three candidate models, then regenerates:

• models/sleep_disorder_pipeline.joblib
• reports/model_comparison.png

Optional paths can be supplied with --data, --model-output, and --report-output.

Predict

Run the interactive predictor:

python -m src.predict

For a repeatable non-interactive prediction, create patient.json:

{
  "Gender": "Male",
  "Age": 31,
  "Occupation": "Software Engineer",
  "Sleep Duration": 7.2,
  "Quality of Sleep": 8,
  "Physical Activity Level": 60,
  "Stress Level": 4,
  "BMI Category": "Normal",
  "Blood Pressure": "120/80",
  "Heart Rate": 70,
  "Daily Steps": 8000
}

Then run:

python -m src.predict --input-json patient.json

Only load .joblib files from sources you trust. Joblib artifacts can execute code when loaded.

Tests

python -m unittest discover -s tests -v

GitHub Actions runs the same test command with Python 3.11 for every push and pull request.

Docker

Build the inference image:

docker build -t sleep-disorder-classifier .

Run it interactively:

docker run --rm -it sleep-disorder-classifier

Or provide a JSON file:

docker run --rm \
  -v "$PWD/patient.json:/tmp/patient.json:ro" \
  sleep-disorder-classifier \
  --input-json /tmp/patient.json

Tech Stack

• Python
• pandas and NumPy
• scikit-learn
• SciPy
• Matplotlib and seaborn
• joblib

Exact tested dependency versions are pinned in requirements.txt.

Limitations

• The dataset is small and synthetic, so reported scores may not generalize to real clinical populations.
• The model has not been clinically validated.
• Predictions depend on self-reported and simplified health features.
• Fold standard deviations describe variation among five correlated validation folds; they are not population confidence intervals or guarantees for future data.
• Model selection and reporting use the same internal cross-validation, so the comparison can be optimistic and should be confirmed on external data.

License

The source code in this repository is licensed under the MIT License.

The dataset is a separate work distributed by its publisher under CC0: Public Domain. The MIT License does not apply to the dataset.

Citation

Do, Chi Khoa (2026). Sleep Health and Lifestyle Dataset Analysis and Prediction.

Contact

Chi Khoa Do — dochikhoa2006@gmail.com