Hybrid Speech Feature Learning Ensembled with Dual Stage Hierarchical Boosting for Classification of Parkinson’s Disease

Abstract

As a long-term and progressive neurodegenerative disease, Parkinson’s disease (PD) is challenging to diagnose at premature stages. Building machine learning models based on speech data for classifying PD objects has proven to be an effective approach. However, PD speech data has highly correlated features, including noise and redundancy, which affect the performance of machine learning models. Although dimensionality reduction algorithms (feature selection and feature extraction) can solve these issues to some extent, it has their limitations. Feature selection algorithms select a subset of the original feature space, thereby losing substantial information by eliminating the features. Feature extraction methods transform the features in a new feature space without reducing the number of features; However, the algorithms suffer from the problem of high aliasing and instability while transforming the high-dimensional data. To address these issues, we propose the HBD-SFDSB Model that leverages hybrid speech feature learning and dual-stage hierarchical Boosting approaches for Accurate PD diagnosis. The main goals of HBD-SFDSB are as follows: (1) constructing a deep chain-based sample space (DCS) by an iterative extraction method. (2) developing a dual-stage feature learning by introducing a hybrid feature learning technique. (3) Performing the dual-stage feature learning recurrently employing the AdaBoost technique to acquire high-quality speech features that significantly enhance the classification accuracy of the machine learning model. The performance of the HBD-SFDSB was evaluated using a self-collected and two publicly available speech datasets. The results show that HBD-SFDSB improves classification accuracy up to 38.71\% compared to available feature reduction and deep learning algorithms. Consequently, HBD-SFDSB may provide valuable insights for future research in this field.