Y. H. Chen
G. Uswatte
K. D. Reilly
L. Hobbs

IMAR, A Data Mining System for Identifying Movement from Accelerometer Recordings

Proc. 2005 Huntsville Simulation Conference, HSC 2005, 1p.
(Publication (Hard Copy: 2006; CD: 2005).


Abstract

In the rehabilitation field it is believed that functional activity in the life situation is the most important outcome to pursue and measure. Development of new cognitive techniques that effectively transfer results obtained in a clinic setting to real-life ones requires innovations in measuring techniques, and more, as we now present. Accelerometers have been used for some to measure bodily movement (arm, leg, chest etc.) involved in the activities of daily living (ADL). Several modeling methods are employed in research, e.g., classification models to recognize types of overall physical activity from accelerometer along. Here we outline several additional modeling efforts, including (limited) statistical analysis, neural networks (NNs), support vector machines (SVMs) and linear classification methods. In our context, the models typically have input consisting of instances of four raw and four reduced data items (thresholded raw data). Output, meanwhile, consists of a 4-level observation measure (taken from scoring videos of subjects during activities) and a 2-level reduced data item. The results show that neural network classification with a time window gives the best performance among all the methods we've used. The time window scheme is indicatory of instance variables requiring context (of other variables' values). This scheme is robust to the extent that even in (limited) studies of individual variations it has held up. In some cases the model input is the raw data; in others, it's the reduced data. This also applies to the output. (Recall that the reduced data is derived by thresholding raw data.) Next stage efforts can be projected. We see expansion out from this to new data and, in further analysis and modeling work: coupling neural network with clustering and classification based schemes onward toward a richer data mining system. Overall, this work reflects themes of biological systems (both real-world data and bio-inspired models), model choices (NNs, SVMs, and so forth), and a range of work from incipient projects to ones which have achieved a degree of maturity.

Key Words:

Data mining, identifying movement, accelerometers.