Comparing algorithms for assessing upper limb use with inertial measurement units

The various existing measures to quantify upper limb use from wrist-worn inertial measurement units (IMU) can be grouped into three categories: (a) Thresholded activity counting, (b) Gross movement score and (c) machine learning. While machine learning algorithms are a promising approach to detect upper limb use, there is currently no knowledge of the information used by these methods, and the data-related factors that influence their performance. A comparison of existing methods was carried out using data from a previous study which was collected from 10 unimpaired and 5 hemiparetic subjects, with annotation to identify periods of functional and non-functional upper limb use. Intra-subject random forest machine learning measures were found to classify upper limb use more accurately than other measures. The random forest measure uses information about the orientation and the amount of movement of the forearm to detect upper limb use. The types of movements and the proportion of functional data included in training/testing set influences the performance of machine learning measures. This study presents the first step towards understanding and optimizing machine learning methods for upper limb use assessment using wearable sensors.