Machine learning and statistical approaches to assessing gait patterns of younger and older healthy adults climbing stairs

• Main Authors: Chan, Herman, Mingjing Yang, Huiru Zheng, Haiying Wang, Sterritt, R., McClean, S., Mayagoitia, Ruth E.
• Format: Conference Proceeding
• Language: English
• Subjects: Accelerometers; Accuracy; activity recognition; classification; Feature extraction; gait patterns; Legged locomotion; Machine learning; Predictive models; statistical analysis; Support vector machines
• ISSN: 2157-9555
• DOI: 10.1109/ICNC.2011.6022097

The following study explores the methods for activity recognition of younger and older adults climbing stairs. There is a correlation to health and the level of activity of an individual, which has captured interest in this field in computing science to determine the level of activity of an individual. The focus of the study is the classification of younger and older gait patterns climbing up and down a set of 13 stairs. From using acceleration data from an accelerometer placed on the lumbro-sacral joint; 14 gait features are extracted and analysed. The machine learning algorithms focused in this study are Multilayer Perceptron (MLP), KStar and Support Vector Machine (SVM). An evaluation of performance among three machine learning algorithms was carried out. MLP was found to provide the highest in accuracy for classification. Accuracy of 95.7% was found for classifying a subject walking either up or down the stairs and an accuracy of 80.6% for classifying whether the subject was younger or older. An evaluation of individual features showed poor performance of classification for young and older subjects climbing up and down stairs, and at most cases failed to distinguish between the two classes. However, subsets of features were created using a sequential feature selection algorithm based on feature ranking and individual feature performance. The performance of each subset was recorded and a subset of the top four features achieved an accuracy of 81.7% for classification between young and older subjects. In comparison, 13 features were required to obtain the best performance of 95.7% to distinguish between up and down classes.