Evaluation of Vertical Ground Reaction Forces Pattern Visualization in Neurodegenerative Diseases Identification Using Deep Learning and Recurrence Plot Image Feature Extraction

Abstract

<jats:p>To diagnose neurodegenerative diseases (NDDs), physicians have been clinically evaluating symptoms. However, these symptoms are not very dependable—particularly in the early stages of the diseases. This study has therefore proposed a novel classification algorithm that uses a deep learning approach to classify NDDs based on the recurrence plot of gait vertical ground reaction force (vGRF) data. The irregular gait patterns of NDDs exhibited by vGRF data can indicate different variations of force patterns compared with healthy controls (HC). The classification algorithm in this study comprises three processes: a preprocessing, feature transformation and classification. In the preprocessing process, the 5-min vGRF data divided into 10-s successive time windows. In the feature transformation process, the time-domain vGRF data are modified into an image using a recurrence plot. The total recurrence plots are 1312 plots for HC (16 subjects), 1066 plots for ALS (13 patients), 1230 plots for PD (15 patients) and 1640 plots for HD (20 subjects). The principal component analysis (PCA) is used in this stage for feature enhancement. Lastly, the convolutional neural network (CNN), as a deep learning classifier, is employed in the classification process and evaluated using the leave-one-out cross-validation (LOOCV). Gait data from HC subjects and patients with amyotrophic lateral sclerosis (ALS), Huntington’s disease (HD) and Parkinson’s disease (PD) obtained from the PhysioNet Gait Dynamics in Neurodegenerative disease were used to validate the proposed algorithm. The experimental results included two-class and multiclass classifications. In the two-class classification, the results included classification of the NDD and the HC groups and classification among the NDDs. The classification accuracy for (HC vs. ALS), (HC vs. HD), (HC vs. PD), (ALS vs. PD), (ALS vs. HD), (PD vs. HD) and (NDDs vs. HC) were 100%, 98.41%, 100%, 95.95%, 100%, 97.25% and 98.91%, respectively. In the multiclass classification, a four-class gait classification among HC, ALS, PD and HD was conducted and the classification accuracy of HC, ALS, PD and HD were 98.99%, 98.32%, 97.41% and 96.74%, respectively. The proposed method can achieve high accuracy compare to the existing results, but with shorter length of input signal (Input of existing literature using the same database is 5-min gait signal, but the proposed method only needs 10-s gait signal).</jats:p>

Development of a Neurodegenerative Disease Gait Classification Algorithm Using Multiscale Sample Entropy and Machine Learning Classifiers

“…For the classification of the HC group and each disease from NDD groups, the proposed algorithm outperforms or equal to the performance to that of the [ 11 , 12 , 13 , 14 , 15 , 16 ]. For classification of any two disease groups from NDD groups, the performance of this study outperforms that of the [ 11 , 12 , 13 , 16 ] but little less than [ 14 ]. However, the accuracy is less than 0.5%.…”

“…The main contribution of this study can be found by comparing the existing literature using the same database [ 9 ]. Table 9 reveals the classification results of the proposed algorithm comparing to other literature [ 11 , 12 , 13 , 14 , 15 , 16 ], the time window length of 10-sec with the KNN model from the proposed algorithm are used to compare with other literature. For the classification of the HC group and each disease from NDD groups, the proposed algorithm outperforms or equal to the performance to that of the [ 11 , 12 , 13 , 14 , 15 , 16 ].…”

“…Table 9 reveals the classification results of the proposed algorithm comparing to other literature [ 11 , 12 , 13 , 14 , 15 , 16 ], the time window length of 10-sec with the KNN model from the proposed algorithm are used to compare with other literature. For the classification of the HC group and each disease from NDD groups, the proposed algorithm outperforms or equal to the performance to that of the [ 11 , 12 , 13 , 14 , 15 , 16 ]. For classification of any two disease groups from NDD groups, the performance of this study outperforms that of the [ 11 , 12 , 13 , 16 ] but little less than [ 14 ].…”

“…The method of dividing the data using overlapping windows was adopted by the definition in [ 16 ] and the number of GF signals samples obtained from the process can be depicted in Equation (5) where l (sec) is the time length of the signal, TW is the time window length (10/20/30/60-sec), d (sec) is the overlapping between consecutive windows, and T is the total subjects in each group.…”

“…Besides, machine learning and deep learning models commonly used in classifying NDD data are support vector machine (SVM) [ 11 , 12 , 13 , 14 , 15 ] and also have other models as k -nearest neighbors (KNN) [ 13 ], multi-layer perceptron (MLP) [ 13 ], and probabilistic neural network (PNN) [ 14 ], least squares SVM (LS-SVM) [ 15 ]. In addition to the machine learning algorithm, state of the art convolutional neural network (CNN) is used for NDD gait classification [ 16 ]. Regarding the cross-validation, leaving one out cross-validation (LOOCV) approach was often used to validate the training process.…”