On the data analysis for classification of elementary upper limb movements

Abstract

Purpose. Body worn inertial sensors could be used to assess rehabilitation of patients with impaired upper limb motor control by detecting and classifying how many times particular arm movements (exercises) are made during normal activities. We present a systematic exploration to determine such a system.
Methods. Kinematic data was collected from 18 healthy subjects using tri-axial inertial sensors (accelerometers and gyroscopes) located at two positions on the dominant arm as four fundamental arm movements were repeated 20 times each. Ten time domain features were extracted from individual and combinations of sensor axes data, and were used to train a classifier. Three different classifiers were investigated: linear discriminant analysis (LDA), quadratic discriminant analysis (QDA) and support vector machine (SVM). Each was verified using a leave-one-subject-out technique for a generalized classification model, and a ten-fold cross validation technique for a personalized classification model.
Results. LDA repeatedly gave the better results when using features extracted from individual sensor axes data. When a personalized learning model is used with LDA, only a single tri-axial sensor (accelerometer or gyroscope) is required to classify all four of the upper limb movements with a sensitivity in the range 92-100%, using as few as 6-10 time-domain features. By comparison, the generalized model using LDA exhibited lower sensitivity and generally required more features (12-18), reflecting the greater variability inherent in a training set comprised of more than one individual’s data.
Conclusions. We demonstrate that body worn inertial sensors can classify elementary arm movements using a low complexity algorithm.

More information

Identifiers

ORCID for Nick Harris: orcid.org/0000-0003-4122-2219

Catalogue record

Export record

Altmetrics