To distinguish ataxic from normal gait using data from full-body motion capture device. To compare accuracy of various machine learning classifiers. To find the most suitable sensor location for ataxia classification.

Group of 7 patients with gait ataxia in spinocerebellar type of multiple sclerosis and 7 healthy volunteers were recorded by “Perception Neuron“ while walking. Each recording was divided into several shorter (20s) parts counting 139 segments in total. Power spectral density of accelerometric data was classified by Support Vector Machine, Bayesian Analysis, K-Nearest Neighbours and Neural Networks. Accuraccy for different sensor locations were compared.

Best accuracy for different sensors locations was achieved on head, neck and shoulders (up to 99.6%). Accuracy of different classifiers in these locations were comparable with maximal difference 1.8%. The worst results were on feet (65-88%, depending on the classification method) and shanks (76-86%).

The evaluation of ataxic walking using motion sensors is a sensitive method, which depends on the location of the sensors. The best accuracy can be achieved by the sensors which are the most distant from the ground. The most likely explanation is the increase of the signal / noise ratio. Noise is comparable for all sensors, while the amplitude of compensatory oscillations increases with distance from the ground.