Development of a Deep Learning Gait Classification Algorithmic Using Time-Frequency Features and Its Verification on Neuro-Degenerative Diseases’ Gait Classification

• Main Authors: FebryanSetiawan, 賴誠信
• Other Authors: Che-Wei Lin
• Format: Others
• Language: en_US
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/4jzevb

碩士 === 國立成功大學 === 生物醫學工程學系 === 107 === Neuro-degenerative diseases (NDDs), such as Amyotrophic Lateral Sclerosis (ALS), Huntington’s Disease (HD), and Parkinson’s Disease (PD), may cause serious gait abnormalities. A comparison of gait abnormalities among healthy controls and NDD subjects will indicate different force pattern variations since the gait force signals are irregular in NDDs. A detection algorithm using a convolutional neural network (CNN) has been developed in this research to classify NDDs based on the gait force signal. The main purpose of this research is to help a physician with screening for NDDs for early diagnosis, efficient treatment planning, and monitoring of disease progression. The database used in this study consisted 64 recordings (five-minutes in each recording) of gait force signals acquired from 16 healthy controls, 13 ALS, 20 HD, and 15 PD subjects. The proposed detection algorithm consists of a windowing process, a feature transformation process, and a classification process. In the windowing process, the five-minute gait force signal was divided into 10, 30, and 60 seconds of successive time windows. There are two feature transformation process compared in this study. In the first one, time domain gait force signal of right or left foot is transformed into a time-frequency spectrogram using a continuous wavelet transform (CWT) spectrogram, specifically a Morlet or Gabor wavelet. In the second apporach, time domain gait force signals from right and left feet are used to compute the wavelet coherence spectrogram. Then, the feature extraction of the time-frequency spectrogram is utilized using a principal component analysis (PCA). The difference force pattern variations among healthy controls, ALS, HD, and PD patients can be distinctly observed from the feature extracted spectrogram images. Finally, CNN is employed in the classification process of the proposed detection algorithm and evaluated using the leave-one-out cross-validation (LOOCV) and the k-fold cross-validation (kfoldCV). As the result, using LOOCV, the highest performance accuracy in the ALS vs. healthy controls classification is 100%, in the HD vs.s healthy controls is 100%, in the PD vs.s healthy controls is 97.42%, in ALS vs. HD is 98.38%, in ALS vs. PD is 100%, in HD vs. PD is 97.90%, and in NDD (ALS+HD+PD) vs. healthy controls is 98.44%. We compared our highest performance accuracy in terms of three classification tasks (ALS vs. healthy control, HD vs. healthy controls, and PD vs. healthy controls) with several existing studies using the same database: Zeng and Wang (2015) [1], Ren et al. (2016) [2], Zhao et al. (2018) [3], and Pham (2017) [4], and find that the proposed method outperforms the performance results in the existing literature. In conclusion, the proposed detection algorithm can effectively differentiate the gait patterns based on a time-frequency spectrogram of a gait force signal between healthy control subjects and patients with neuro-degenerative diseases.