Parkinson's Disease Detection Using Machine Learning

⚠️ DISCLAIMER: This project is for educational and research purposes only. It is NOT a medical diagnostic tool. Do not use it to make health decisions. Always consult a qualified healthcare professional for medical advice.

Detects Parkinson's disease from voice recordings using an SVM (Support Vector Machine) classifier. The system analyzes phonation features — the sounds produced when pronouncing sustained vowels — since phonation is the most affected component of speech in Parkinson's patients.

How It Works

1. Feature Extraction: Voice recordings (.wav) are processed using PRAAT (via parselmouth) to extract acoustic features: jitter, shimmer, NHR, HNR, pitch statistics, pulse counts, and period measurements.
2. Training: A Gradient Boosting classifier (200 estimators, max_depth=5) is trained on the UCI voice dataset (1040 recordings from 56 subjects). Feature scaling is applied via StandardScaler, and the scaler is saved alongside the model.
3. Prediction: New voice samples are analyzed, features are extracted and scaled using the saved scaler, then classified as Parkinson's-positive or healthy.

Project Structure

├── features.py           # Shared feature extraction module (PRAAT parsing + validation)
├── train.py              # Train the SVM model and save it with the scaler
├── predict.py            # CLI prediction from a .wav file
├── gui.py                # Tkinter GUI for file selection and prediction
├── final2.csv            # Training dataset (1040 samples, 29 columns)
├── requirements.txt      # Pinned Python dependencies
├── LICENSE               # MIT License
└── README.md

Setup

# Python 3.10+ required
pip install -r requirements.txt

Usage

1. Train the model

python train.py

Example output:

Cross-validation accuracy: 0.6670 (+/- 0.0567)

Confusion Matrix:
[[74 33]
 [26 75]]

              precision    recall  f1-score   support

     Healthy       0.74      0.69      0.71       107
  Parkinsons       0.69      0.74      0.72       101

    accuracy                           0.72       208

Model and scaler saved to svmclassifier.pkl

2. Predict from a .wav file (CLI)

python predict.py path/to/voice_recording.wav

Healthy — No Parkinson's Detected

3. Predict using the GUI

python gui.py

Use File → Open to select a .wav file containing a sustained vowel ('a' or 'o'), then click Detect.

Dataset

• Source: UCI Machine Learning Repository — Parkinson Speech Dataset with Multiple Types of Sound Recordings
• License: CC BY 4.0
• Samples: 1040 recordings from 56 subjects (sustained vowels 'a' and 'o', 3 times each)
• Features: 26 acoustic features including jitter variants, shimmer variants, autocorrelation, noise-to-harmonics ratio, pitch statistics, pulse/period measurements, voicing breaks
• Label: Binary (1 = Parkinson's, 0 = Healthy) at column index 28

Citation

If you use this dataset, please cite:

Sakar, B.E., Isenkul, M.E., Sakar, C.O., Sertbas, A., Gurgen, F., Delil, S., Apaydin, H., Kursun, O. (2013). Collection and Analysis of a Parkinson Speech Dataset with Multiple Types of Sound Recordings. IEEE Journal of Biomedical and Health Informatics, 17(4), 828-834.

Model Accuracy

The model achieves ~72% accuracy on final2.csv using Gradient Boosting with all 26 acoustic features. This is an improvement over the original SVM with 10 PSO-selected features (~64%), achieved by using all available PRAAT features (including voicing break metrics) and hyperparameter tuning via GridSearchCV.

Limitations

• No automated tests. Contributions adding unit tests for extract_features(), train(), and predict() are welcome.
• Model accuracy is ~72% on the available dataset. See Model Accuracy for context.
• Pickle security: The model is saved/loaded via pickle, which can execute arbitrary code. Only load .pkl files you generated yourself with train.py. Never load untrusted pickle files.

License

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

Contributing

Contributions are welcome. See CONTRIBUTING.md for guidelines.