Gait Analysis for Severity Rating of Parkinson’s Disease Using Machine Learning and Deep Learning Algorithms

Abstract

Parkinson’s disease (PD) is a chronic neurodegenerative brain disorder, which affects the ability of the person to perform the regular activities. While diagnosing PD, neurologists often use several clinical manifestations like motor and non-motor symptoms and rate the severity based on the Unified Parkinson Disease Rating Scale (UPDRS). This kind of rating largely depends on the expertise and experience of the clinicians. and it is not only subjective but also inefficient. Gait pattern, which plays a vital role in assessing the human mobility, is a significant biomarker to classify whether the subject is healthy or affected with PD. Hence, in this work, we aim to investigate the gait pattern of healthy and PD subjects using machine learning (ML) and deep learning (DL) algorithms to design an automatic and non-invasive PD diagnosis and severity rating system that can assist the neurologists in their daily PD diagnosis. In this regard, the vertical ground reaction force (VGRF) gait dataset is utilized and the following three ML and DL classification algorithms are explored: 1.Supervised machine learning based classifier 2. Convolutional neural network (CNN) classifier and 3. Long-short term memory(LSTM) network classifier. Firstly, To obtain the optimal feature set for the classification model, a correlation based feature selection technique is employed. Secondly, the statistical analysis of the VGRF sensor data is performed to differentiate between healthy control and PD patients. Four supervised machine learning algorithms namely decision tree(DT), support vector machine (SVM), ensemble classifier(EC) and Bayes classifier(BC) are used to classify the stages of PD based on Hoehn and Yahr (H&Y) scale. Moreover, to avoid data overfitting problem and enhance the classification accuracy, the 10 fold cross validation technique is utilized. Even though the supervised machine learning classifiers yield high accuracy in detecting the PD, one of the major limitations is the need for hand-crafted features. iv Hence, to overcome the hand-crafted feature approach, secondly, we explore the CNN classifier for multi-class classification for various frame sizes. To avoid the data over-fitting, L2 regularization technique, which penalizes the weight parameters of the nodes, is used in combination with the dropout layer. For optimizing the loss function, a stochastic gradient descent (SGD) algorithm is employed because it reduces the computational burden for large dataset. Experimental results substantiate that the proposed DCNN architecture outperforms state-of-the-art artificial neural network (ANN) classifiers and achieves the highest classification accuracy of 98.45%. Thirdly, to evaluate the potential of LSTM classifier, which is highly suitable for learning the long-term temporal dependencies in the gait cycle, we study the PD rating based on modified H&Y scale and UPDRS scale for three different walking scenarios. For solving the cost function, Adam, a stochastic gradient-based optimizer, is employed and the severity of PD is categorized based on UPDRS and H&Y scale. The experimental results reveal that Adam optimized LSTM network can effectively learn the gait kinematic features and offer an average accuracy of 98.6% for binary classification and 96.6% for multi-class classification