A deep learning-based approach to detect Parkinson’s disease through speech analysis

Abstract

Parkinson’s Disease (PD) is a progressive type of neurodegenerative disorder in which timely diagnosis is extremely important in managing the disease, yet it is difficult because of the insidious nature of the symptoms. The conventional diagnostic approaches usually use subjective clinical evaluation which might fail to reveal early biomarkers. This study fills this gap by proposing a deep learning-based framework that can be used to detect PD non-invasively with speech signals. In this paper, we present a multi-model fusion architecture that makes use of two different representations of one audio signal, which are Mel-spectrograms to represent spectral-visual features and Wav2Vec 2.0 embeddings to represent high-level contextual phonetic features. We extractedWav2Vec2 embedding features from a fully connected neural network that was trained on the Wav2Vec2 embeddings. We then ran the melspectrograms though CNN and other convolutional networks such as ResNet18, ResNet50, Inception v3, Efficient Net B0, and Efficient Net B4. We then extracted melspectrogram features from all these convolutional networks and concatenated them seperately with the Wav2Vec2 embedding features. The cocnatenated features then went through a commonly shared final classifier. One of the most important contributions of this work is the introduction of a new rational sine activation function that gives the model a better capability of learning nonlinear decision boundaries. The results of the experiment conducted on the Italian Parkinson voice and speech data show that the proposed fusion model attains a test accuracy of 85.29% which is higher than what the other architectures that involve transfer learning models such as ResNet18, ResNet50, Inception v3, Efficient Net B0, and Efficient Net B4 have achieved. Moreover, a strict cross-corpus test on the unseen Vowel dataset shows that although domain shift is a universal problem, the suggested fusion architecture demonstrates better generalization with respect to the baseline models, which are adversely affected by the problem. The paper has validated the hypothesis that a combination of mixed signal representations with mathematically optimized activation functions is a promising avenue towards automated and objective screening of PD.