Decoding of Deep Learning of Parkinson's Disease Dysphonia Based on Partial Least Squares Regression

Speech signals have recently gained a lot of attention in Parkinson's Disease (PD) as these signals can be used to distinguish whether the subjects have the disease or not. Some studies classified patients with PD and healthy people based on acoustic features to make the diagnosis explicable. Other studies distinguished these through deep learning features which had an excellent performance. This paper combines the advantages of the two methods and proposes a method that decoding the deep learning features from the acoustic features based on Partial Least Squares Regression (PLSR). In this paper, we construct a framework that decode the Deep Learning features from the acoustic features of speech signals. First, deep learning features learned using convolutional neural networks and Mel-frequency cepstral coefficients (MFCC) as one of the acoustic features are extracted. Next, PLSR is utilized to obtain the multivariate mapping model between the MFCC and the deep learning features. Finally, the performance of the proposed method is validated through two different datasets, namely the Chinese Pronunciation Parkinson Detection Dataset (CPPDD), and the Sakar's Parkinson Detection Dataset (SPDD). The experimental results show that the decoding accuracy of the proposed method on CPPDD and SPDD is 80.43% and 98.66%, respectively. This indicates that the proposed method that decodes the deep learning features from the acoustic features based on PLSR is feasible from the aspect of performance.