High-resolution superlet transform based techniques for Parkinson's disease detection using speech signal

Abstract

Parkinson's Disease (PD) is a matter of great concern when it comes to the health management of elderly people. Tremors, muscle stiffness, change in cognitive abilities, and dysarthria are some of the frequently encountered symptoms, commonly found in patients diagnosed with PD. A substantive progression of PD can be seen in the patient's body, caused by the degeneration of the dopamine-generating cells. PD is still a non-curable disease but its progression can be slowed down with some medications and therapies. Thus, early detection of the disease becomes a necessity for much-needed care of the patients. Recent developments in the technology sector have rewarded us with an opportunity for early detection of PD using speech signals. In this paper, a Deep Neural Network (DNN) based approach using spectrograms of speech signals generated by Superlet Transform (SLT) is proposed for the detection of PD. The SLT converts the 1-D speech signals into 2-D spectrogram. These spectrograms of the speech signal are then applied to different DNN classifiers namely InceptionResNetV2, VGG-16, and ResNet50v2 for PD detection. The proposed method is evaluated on sustainable vowels, modulated vowels, DDK analysis, and isolated words of PC-GITA dataset and vowels of ItalianPVS dataset. Several evaluation measures have been used to estimate the robustness and explainability of this work. The developed framework gives the best performance on modulated vowels using VGG-16 on PC-GITA dataset. The VGG-16 with SLT achieved an overall accuracy of 92% with sensitivity, specificity, precision, and F-1 score of 92%, 91%, 95%, and 93%, respectively. The proposed method achieved an accuracy of 96% on the ItalianPVS dataset. This method outperforms the state-of-the-art methods like Hilbert spectrum analysis, Empirical Mode Decomposition (EMD), Continuous Wavelet Transform (CWT), and Short-Time Fourier Transform (STFT) for PD detection using different speech data sources.