Multimodal Parkinson's Disease Detection using Random Forest

Overview

This repository provides a machine learning model for Parkinson's Disease detection using a multimodal approach that combines speech-based acoustic biomarkers and hand-drawn image features.

The model integrates clinically relevant voice features with Histogram of Oriented Gradients (HOG) extracted from spiral and wave drawings to improve diagnostic performance.

The classifier is a Grid Search optimized Random Forest model trained on fused multimodal features.

System Architecture

Voice Recording
        │
        ▼
 Voice Feature Extraction
        │
        ├──────────────┐
        │              │
        ▼              ▼
Drawing Image     HOG Feature Extraction
        │              │
        └──────┬───────┘
               ▼
      Feature Concatenation
               ▼
 Random Forest (Grid Search)
               ▼
      Parkinson Prediction

Problem Statement

Parkinson's Disease is a progressive neurological disorder where early diagnosis is essential for effective treatment.

Traditional diagnosis often depends on clinical examination. This project demonstrates how machine learning can assist clinicians by analyzing multiple patient modalities simultaneously.

Model Details

Property	Value
Model	Random Forest Classifier
Optimization	Grid Search CV
Task	Binary Classification
Framework	Scikit-learn
Input	Voice + Drawing Features
Output	Healthy / Parkinson's Disease

Dataset

Voice Dataset

Source: UCI Parkinson's Dataset
Samples: 195
Parkinson's: 147
Healthy: 48

Drawing Dataset

Spiral and Wave Drawing Dataset

Total Images: 207
Training Images: 147
Testing Images: 60

Multimodal Dataset

Voice and drawing features were combined into a single feature vector after preprocessing and class balancing using Random Oversampling / SMOTE.

Input Features

Voice Features

Fundamental Frequency (Fo)
Highest Frequency (Fhi)
Lowest Frequency (Flo)
Jitter
Shimmer
NHR
HNR
RPDE
DFA

Drawing Features

Histogram of Oriented Gradients (HOG)

Preprocessing includes:

Grayscale conversion
Image resizing (250×250)
Otsu Thresholding
HOG Feature Extraction

Feature Fusion

The multimodal feature vector is generated by concatenating the processed voice features and HOG image descriptors.

model_input = np.concatenate((voice_features, img_features), axis=1)

Performance

Metric	Score
Accuracy	92.73%
Precision	100.00%
Recall	90.70%
F1 Score	95.12%

Confusion Matrix

	Predicted Healthy	Predicted Parkinson's
Actual Healthy	12	0
Actual Parkinson's	4	39

Installation

git clone https://github.com/yourusername/multimodal-parkinsons-random-forest.git

cd multimodal-parkinsons-random-forest

pip install -r requirements.txt

Repository Structure

multimodal-parkinsons-random-forest/

├── README.md
├── parkinson_multimodal_random_forest.pkl
├── requirements.txt
├── LICENSE
├── src/
├── examples/
└── images/

Intended Use

This model is intended for:

Educational purposes
Academic research
Machine Learning experimentation
Healthcare AI demonstrations

It is not intended for clinical diagnosis or medical decision-making.

Limitations

Dataset size is relatively small.
Performance depends on the quality of voice recordings and drawing images.
The model has not been clinically validated.
Predictions should not replace professional medical evaluation.

Ethical Considerations

This project is developed solely for research and educational purposes.

Medical AI systems should always be validated by healthcare professionals before being used in real-world clinical settings.

License

MIT License

Author

Sarthak.

AI/ML Engineer

Specializing in Machine Learning, Computer Vision, NLP, LLMs, and Generative AI.

Downloads last month: -; Downloads are not tracked for this model. How to track