Discrete wavelet transform based data representation in deep neural network for gait abnormality detection

• Published in: Biomedical signal processing and control Vol. 62; p. 102076
• Main Authors: Chakraborty, Jayeeta, Nandy, Anup
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.09.2020
• Subjects: Cerebral palsy gait; Convolutional neural network; Deep learning; Discrete wavelet transform; Gait pattern classification; Wearable inertial sensor; Deep learning; Gait pattern classification; Discrete wavelet transform; Cerebral palsy gait; Convolutional neural network; Wearable inertial sensor
• ISSN: 1746-8094; 1746-8108
• DOI: 10.1016/j.bspc.2020.102076

•A novel approach for classification of abnormal gait pattern using inertial sensors data.•Auto-correlation technique is used for segmentation of signals.•Discrete wavelet transformation is used for data representation.•Deep neural network architecture modeling is utilized for automatic feature extraction.•The proposed method has achieved significant improvement in performance over the state-of-the-art deep learning techniques for inertial sensors. Detection of abnormal gait patterns using wearable sensors remains a major challenge in clinical gait analysis and rehabilitation field. Despite the success of recent researches using deep learning techniques, the prospects of improvement in the classification process with the help of modification in data representation is largely overlooked. In this paper, a deep neural network-based framework is proposed where discrete wavelet decomposition is used for data representation to detect abnormal gait patterns using inertial sensors. In the proposed approach, the walking gait data of healthy children and cerebral palsy children are collected using two inertial sensors. Discrete wavelet transform is applied to signal segments to form decomposed signal segments. A multi-channel 1-dimensional convolutional neural network (1D-CNN) model is trained with the decomposed signals. The proposed method achieves 96.4% and 90.97% accuracy for segment-wise and subject-wise evaluation respectively. The performance of the proposed model is compared with the state-of-the-art methods as well as with a basic 1D-CNN model trained with signals directly. Analysis of the result shows that the proposed method performs significantly better than the basic CNN model and also exceeds over the performance of the state-of-the-art methods. An investigation is done on the effect on the performance of the model with varying levels of wavelet decomposition which reveals that at level 2, the proposed method reaches the highest accuracy and lowest loss value. When tested with 100 random samples, the wavelet representation generates higher area-under-curve scores for deep learning based techniques, compared to empirical mode decomposition representation method.