Parkinson's disease detection and stage classification: quantitative gait evaluation through variational mode decomposition and DCNN architecture

• Published in: Connection science Vol. 36; no. 1
• Main Authors: E, Balaji, Elumalai, Vinodh Kumar, Sandhiya, Dhanasekaran, Swarna Priya, R. M., Shantharajah, S. P.
• Format: Journal Article
• Language: English
• Published: Abingdon Taylor & Francis 31.12.2024; Taylor & Francis Ltd
• Subjects: Artificial neural networks; Biomarkers; Classification; Cognition; Decomposition reactions; deep convolutional neural network; Gait; gait analysis; Parameter identification; Parkinson's disease; recurrence plot; Regularization; Spectrum analysis; variational mode decomposition; Vertical forces
• ISSN: 0954-0091; 1360-0494; 1360-0494
• DOI: 10.1080/09540091.2024.2383894

Parkinson's disease (PD) is a progressive, debilitating neurological movement disorder that affects the person's muscle control, movement, speech, cognition and dexterity. For diagnosing PD in a clinical setting, in addition to the neurological examinations, clinicians use the unified Parkinson disease rating scale (UPDRS) to assess the motor and non-motor impairments. Such a clinical assessment highly depends on the experience and expertise of the clinicians, and it may result in biased evaluation. Hence, to assist the clinicians, we put forward a gait analysis-based deep convolutional neural network (DCNN) framework which leverages the potentials of variational mode decomposition (VMD) technique with the recurrence plots (RP) to enhance the PD severity classification performance. Specifically, transforming the VMD modes of vertical ground reaction force (VGRF) time series data into two-dimensional texture images to capture the temporal dependency, this work trains the DCNN classifier through recurrence images for its ability to extract the discriminative features among the PD severity levels. For evaluation, this study utilises the VGRF dataset of 93 PD subjects and 73 healthy controls from Physiobank for three different walking tests. Consequently, utilising VMD, RP and DCNN in a unified framework, this investigation shows that the PD severity rating can be significantly enhanced through DCNN model that is trained using RP of dominant intrinsic mode functions (IMFs). The novelty of the proposed framework lies in identifying the prominent gait biomarkers through dominant IMFs from power spectral analysis for reducing the computational burden of DCNN. Moreover, to handle the data over-fitting issue in the classifier, L2 regularisation technique, which penalises the weight parameters of the nodes, is used in combination with the dropout layer. Experimental results underscore that the proposed VMD-RP-DCNN architecture can address the spectral overlapping issue in VGRF decomposition and achieve an average PD severity prediction accuracy of 98.45%.