Novel Speech-Recognition Approach for the Early Detection of Parkinson's Disease using a Transformer-based Hybrid CNN-GRU Model

Abstract

Parkinson's disease (PD) is a neurodegenerative disease determined by motor and nonmotor symptoms that affect millions of people around the globe due to its neurodegenerative nature. Early detection of PD is very important for timely intervention and improved patient quality of life. Speech-related changes, such as altered prosody and articulation, are common in individuals with PD, making speech analysis a promising method for early diagnosis. In this study, we propose a novel approach for the early detection of Parkinson's disease utilizing a hybrid Convolutional Neural Network combined with a Gated Recurrent Unit and a Transformer model. The hybrid architecture integrates the strengths of different deep learning components to extract and analyze salient features from speech signals. The convolutional neural network (CNN) component captures local patterns and spectral features from short-term speech segments. The Gated Recurrent Unit (GRU) component captures temporal dependencies and long-range contextual information, allowing the model to discern subtle changes in speech dynamics indicative of Parkinson's disease. The Transformer component enhances the model's ability to capture global dependencies and relationships across different speech features, contributing to a more comprehensive understanding of the underlying patterns. The key features of our proposed approach include its ability to process raw speech signals directly, mitigating the need for extensive feature engineering. The hybrid model's adaptability to different speech datasets enhances its generalizability, making it applicable across diverse patient populations. Our experimental results demonstrate superior performance when comparing with other relative and existing deep learning models.