Volume 31, Issue 1, 2022
DOI: 10.24205/03276716.2022.6003
A Layer-Wise Relevance Propagation Based Feature Selection and Hybid Classification Model for Automatic Detection of Parkinson's Disease Using Gait Signals
Abstract
Gait irregularities frequently result from orthopaedic or neurological diseases, and they can have serious repercussions including limiting movement and falling. Gait analysis is essential for tracking gait irregularities, and identifying underlying impairments can aid in creating treatment plans. Spatio-temporal, kinematic, and muscle activation gait features must all be examined in today's multi-modal gait analysis. The second most prevalent neurodegenerative ailment, PD (Parkinson's disease), is brought on by the midbrain's premature neuronal loss. The absence of valid neuropathological criteria prevents a firm pathological diagnosis of PD. To classify PD severity, DLTs (deep learning techniques) are utilised to analyse and combine raw data of gait-induced GRFs (ground reaction force) for PD diseased and healthy patients. The current LSTM model also requires a lot of training time and is very sensitive to random weight initialization. For the effective operation and reliable prediction of PD data, this study introduces an efficient feature selection and hybrid CNNs (Convolution Neural Networks) with LSTM (Long Short Term Memory). In the beginning, processing is done to remove noisy data. Then, to evaluate the results of the models and reveal which elements in the spatiotemporal gait GRFs signals are most important for the models' predictions, feature selection is carried out via LRPs (Layer-wise relevance propagations). This enables their assignment to gait events, suggesting that heel strikes and body balances are best suggestive gait aspects for the categorization of healthy gait. Landing of the foot and body balances are the most affected during late stages of PD. Finally, Hybrid CNNs with LSTM for the reliable prediction of PD data and efficient operation. The suggested models can be helpful for identifying changes in postural balance and grading PD severities since they are robust towards noises and process/ classify big datasets efficiently.
Keywords
Sensor applications, Parkinson's Disease, gait characteristics, Layer-wise relevance propagation (LRP), Hybrid Convolutinal Neural Network (CNNs ) with Long Short Term Memory (LSTM)