A generic optimization and learning framework for Parkinson disease via speech and handwritten records

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Parkinson's disease (PD) diagnosis is challenging due to late-stage symptom identification and symptom overlap with other conditions.
• Early diagnosis and treatment are crucial for managing PD symptoms and slowing progression.
• Current diagnostic methods often lack the sensitivity for early detection.

Purpose of the Study:

• To develop a generic framework for early Parkinson's disease diagnosis using handwritten images and speech signals.
• To evaluate the efficacy of deep learning models and machine learning algorithms in PD detection.
• To introduce a novel feature extraction technique for speech signals in PD diagnosis.

Main Methods:

• Handwriting images: 8 pre-trained Convolutional Neural Networks (CNNs) utilized transfer learning and Aquila Optimizer on the NewHandPD dataset.
• Speech signals: Numerical features extracted using 16 algorithms and fed to 4 machine learning algorithms (Grid Search optimized). Graphical features extracted using 5 techniques and fed to 8 CNNs.
• Novel speech feature extraction: Variable speech-signal-segment-durations used for segmentation, generating 5 datasets with 281 numerical features.

Main Results:

• Handwriting analysis: Achieved a maximum accuracy of 99.75% using the VGG19 CNN structure on the NewHandPD dataset.
• Speech analysis (numerical): Reached 99.94% accuracy using KNN and SVM algorithms with combined numerical features from the MDVR-KCL dataset.
• Speech analysis (graphical): Achieved 100% accuracy using mel-specgram graphical features and the VGG19 CNN structure on the MDVR-KCL dataset.

Conclusions:

• The proposed framework demonstrates high accuracy in diagnosing Parkinson's disease from both handwriting and speech data.
• The novel speech feature extraction technique and deep learning models show significant potential for early and accurate PD detection.
• The results surpass existing state-of-the-art methods, offering a promising tool for clinical application.