RehabMove 2018: Automatic recognition of gait patterns with machine learning

Y. Zhou, T. Hortobágyi, C.J.C. Lamoth
2019 Zenodo  
PURPOSE: The purpose of this study was to automatically identify gait patterns of geriatric patients with and without cognitive impairment using non-linear machine learning. METHODS: Geriatric out-patients with (n=72, age: 80.8 ±5.98) and without cognitive impairment (n=115, age: 79.7 ±5.68) participated in the study. Cognition was assessed by the Mini Mental State Examination. Trunk accelerations were measured with a 3D accelerometer (DynaPort® MiniMod, McRoberts BV and iPod touch 4G, iOS 6,
more » ... ple Inc.; sample frequency ±100 Hz) during three minutes of walking. From the 3D accelerometer signals 23 dynamic gait variables were calculated related to gait pace, stability, regularity, variability and regularity. A Kernel (polynomial and Gaussian kernel functions) Principle Component Analysis (KPCA) was applied to extract underlying gait features and reduce the dimensionality of the data. Thereafter, a non-linear classification Support Vector Machine approach (SVM) was compared with Artificial Neural Networks (ANN). RESULTS: KPCA reduced gait data dimensions efficiently from 23 dimensions to five dimensions, explaining 100 % of the variance, and representing gait features in pace, synchronization, regularity, and variability. Preliminary analyses showed that classification accuracy of SVM (83%) and ANN was similar and both could identify gait of geriatric patients with and without cognitive impairment. However, ANN is sensitive to parameter selection. With regard to SVM, different kernel functions resulted in similar classification performance with less parameter influence. Compared with ANN, SVM performed better with the type of data used in the present study. CONCLUSIONS: Cognitive impairment affects specific gait features that can be identified by non-linear approaches like KPCA in combination with SVM. Based on gait dynamics representing the quality of the gait, participants with and without cognitive impairment could be labeled automatically.
doi:10.5281/zenodo.2607030 fatcat:vcawyyp45bgtpm73v4ymubehsq