Adaptive Myoelectric Human-Machine Interface for Video Games Mohammadreza Asghari Oskoei and Huosheng Hu School of Computer Science and Electronic Engineering, University of Essex Wivenhoe Park, Colchester CO4 3SQ, Essex, United Kingdom Email: masgha@essex.ac.uk ; hhu@essex.ac.uk Abstract – This paper proposes adaptive schemes for myoelectric based human-machine interface (HMI) applied to a video game. Adaptive schemes modify the classification criteria to keep a stable performance in long-term operations. Online support vector machine (SVM) is used as the core of classification to facilitate incremental training during run-time. Supervised and unsupervised methods are individually employed to update online training data set. The experimental results show that the proposed adaptive schemes increase the achieved scores and make a stable performance for myoelectric HMI. Index Terms – Myoelectric HMI, Adaptive Schemes, Video Game, Rehabilitation I. INTRODUCTION The myoelectric human-machine interfaces (HMI) could be employed as an alternative interface in powered wheelchairs and video games for the disabled people [1][2]. Moreover, the manifestation of fatigue in myoelectric signals is perceivable during long-term muscular activities [2]. However, reliability and robustness of a myoelectric HMI is still an open question particularly in a long-term operation. Myoelectric signal (MES) inherently has a complex stochastic structure and its characteristics are intensively dependent to subject’s physical and physiological conditions, muscular activities, and data collection conditions. Existence of these intense dependencies has led us to employ machine learning approaches, such as support vector machine (SVM), in developing myoelectric HMI [3] since they are capable to adapt themselves with signal characteristics using real samples produced before and during the run-time. Training a classifier (e.g. SVM) is an adaption process, in which its parameters are being adjusted using samples generated by a subject that is going to use HMIs and in conditions that are supposed to be for HMI’s work. This is the reason that the training before application, known as offline training, is required. Off-line training discriminates MES patterns corresponding to muscular activities and keeps them in the form of parameters that construct the boundaries between classes. Its period should be adequately enough to collect comprehensive samples that represent different states of muscular activities and the influencing factors on them. In spite of comprehensive offline training, existence of novel samples for myoelectric HMI during run-time is inevitable. This is due to stochastic process of MES generation, slowly growing phenomena having impact on MES such as fatigue, and external factors such as electrode displacement. Hence, myoelectric HMI needs schemes that maintain its performance robust and reliable. A closed loop control system can provide a stable performance using feedbacks. As we know, visual feedback and stimulated sensory signals (toward the body) are two feedbacks that can be adopted in myoelectric HMIs. However, visual feedbacks that continuously involve the mind are not convenient for long-term applications. Meanwhile, the stimulated sensory signal is not always cost effective and practical. For example, when we grab an egg, we do not think about how hard we should grasp after we have gained such experience. Instead, our nervous system automatically takes care of grabbing an egg without breaking or dropping [4]. Adaptive control that involves modifying the control criteria to cope with parameter changes is another option to keep a stable performance for myoelectric HMI. Having a proper model of deviations in MES patterns is a key issue to stabilize the accuracy of manipulating commands. The model has to distinguish regular changes that represent various commands from unwanted changes (i.e. deviation) resulted in accuracy decline. It should differentiate transient states as well, which are highly unpredictable and even contradictory [5], from steady states that carry useful information. Changes in MES patterns are either gradual or significant. The gradual changes can be resolved by adaptive schemes otherwise the system would need a re-configuration. Fatigue is a time-related factor that leads to gradual performance variations. It can be named as the dominant factor that affects steady states of MES in a long-term operation [6]. Online training is the core of adaptive schemes for pattern recognition based HMIs. It rebuilds the boundaries between classes using updated training data set (TDS) during run-time operations. There are two challenges for online training: (i) updating TDS to distinguishing deviations in MES patterns that lead to HMI performance decline, and (ii) online training algorithms that modify the classifier’s parameters in run-time [7]. This paper investigates supervised and unsupervised methods to update TDS, employs online SVM to handle online training smoothly, and evaluates adaptive schemes by applying them to a video game. The rest of this paper is organised as follows. Section II introduces online SVM to manage incremental training in adaptive schemes. Section III describes the proposed supervised and unsupervised adaptive schemes for myoelectric HMI. The experiments conducted to evaluate the adaptive schemes are presented in Section IV, and finally, a conclusion and future work are given in Section V. 1015 978-1-4244-2693-5/09/$25.00 ©2009 IEEE Proceedings of the 2009 IEEE International Conference on Mechatronics and Automation August 9 - 12, Changchun, China