A comprehensive hybrid framework for Parkinson’s disease detection: integrating handcraft features along with deep learning-based feature extraction with variational autoencoder and traditional machine learning techniques for classification

Abstract

Neurodegenerative disorders, such as Parkinson’s disease, present a significant medical challenge, necessitating innovative approaches for detection. This thesis introduces a comprehensive hybrid framework that combines handcrafted features and deep learning techniques to improve the accuracy of Parkinson’s disease detection. The approach leverages pre-trained convolutional neural networks (CNNs) such as VGG16, MobileNet, and EfficientNet, ResNet to extract features of melspectrograms generated from the voice samples. A second contribution is the extraction of handcrafted features from the raw audio data. The features extracted are encoded using a Variational Autoencoder (VAE), which further reduces the dimension and integrated them to further train the machine learning algorithms such as Random Forest Classifier (RFC), K-Nearest Neighbour (KNN), Logistic Regression (LR), Support Vector Machine (SVM), and XGBoost to differentiate. To achieve this, we leveraged the combined strengths of these models by integrating both handcrafted and deep learning features to construct a highly optimized and effective classification model using a hybrid approach that highlights the potential of feature extraction techniques and advanced machine learning algorithms for improving the detection and diagnosis of Parkinson’s disease and facilitating more progress in computational healthcare and early stage diagnostics.