Application of support vector machines to detect hand and wrist gestures using a myoelectric armband
The propose of this study was to assess the feasibility of using support vector machines in analysing myoelectric signals acquired using an off the shelf device, the Myo armband from Thalmic Lab. Background: With the technological advances in sensing human motion, and its potential to drive and control mechanical interfaces remotely or to be used as input interfaces, a multitude of input mechanisms are used to link actions between the human and the robot. In this study we explored the feasibility of using human arm’s myoelectric signals with the aim of identifying a number of gestures automatically. Material and methods: Participants (n = 26) took part in a study with the aim to assess the gesture detection accuracy using myoelectric signals. The Myo armband was used worn on the forearm. The session was divided into three phases, familiarisation: where participants learned how to use the armband, training: when participants reproduced a number of random gestures presented on screen to train our machine learning algorithm; and recognition: when gestures presented on screen were reproduced by participants, and simultaneously recognised using the machine learning routines. Support vector machines were used to train a model using participant training values, and to recognise gestures produced by the same participants. Different Kernel functions and electrode combinations were studied. Also we contrasted different lengths of training values versus different lengths for the recognition samples. Results: One participant did not complete the study due to technical errors during the session. The remaining (n = 25) participants completed the study allowing to calculate individual accuracy for grasp detection. The overall accuracy was 94.9% with data from 8 electrodes , and 72% where only four of the electrodes were used. The linear kernel outperformed the polynomial, and radial basis function. Exploring the number of training samples versus the achieved recognition accuracy, results identified acceptable accuracies (> 90%) for training around 3.5s, and recognising grasp episodes of around 0.2s long. The best recognised grasp was the hand closed (97.6%), followed by cylindrical grasp (96.8%), the lateral grasp (94%) and tripod (92%). Discussions: The recognition accuracy for the grasp performed is similar to our earlier work where a mechatronic device was used to perform, record and recognise these grasps. This is an interesting observation, as our previous effort in aligning the kinematic and biological signals had not found statistically significant links between the two. However, when the outcome of both is used as a label for identification, in this case gesture, it appears that machine learning is able to identify both kinematic and electrophysiological events with similar accuracy. Future work: The current study considers use of support vector machines for identifying human grasps based on myoelectric signals acquired from an off the shelf device. Due to the length of sessions in the experiment, we were only able to gather 5 seconds of training data and at a 50Hz sampling frequency. This provided us with limited amount of training data so we were not able to test shorter training times (< 2.5s). The device is capable of faster sampling, up to 200Hz and our future studies will benefit from this sampling rate and longer training sessions to explore if we can identify gestures using smaller amount of training data. These results allows us to progress to the next stage of work where the Myo armband is used in the context of robot-mediated stroke rehabilitation.