Ensemble classification for robust discrimination of multi-channel, multi-class tongue-movement ear pressure signals Michael Mace, Khondaker Abdullah-Al-Mamun, Shouyan Wang, Lalit Gupta and Ravi Vaidyanathan Abstract— In this paper we introduce a robust classifica- tion framework for tongue-movement ear pressure signals based around an ensemble voting methodology. The ensemble members are comprised of different combinations of sensor inputs i.e. two in-ear microphones and an acoustic gel sensor positioned under the chin of the individual and classification using three different base models. It is shown that by using all nine ensemble members when compared to the individual (base) models, the average misclassification rate can be reduced from 23% to 2.8% when using the majority voting strategy. The correct classification rate is improved from 76% to 92.4% when utilizing either the borda count or condorcet methods. This is achieved through a combination of rejection based on ambiguity in the ensemble and diversity in the misclassified instances across the ensemble members. I. INTRODUCTION Afflictions of the sensory-motor system both physical and neurological can profoundly inhibit human movement. The extremities tend to be at higher risk due to their inherent distance from the brain and body, thus increasing the po- tential for severing of the peripheral nervous system and/or limbs. Upper extremity motor loss can be induced by spinal cord injury (SCI), paraplegia, congenital limb deformities and stroke to name a few. Over the last few decades, a multitude of research has been conducted towards providing novel solutions that replace or compensate these degraded pathways. One particular area of interest is providing an individual with new ways for communicating with assistive technologies. This involves thinking of new and creative methods by which a user can express their intention and thereby control peripheral devices. The use of the head, tongue, eyes and brain in providing these new communi- cation pathways, have been employed by researchers, due to the robust functionality of these craniofacial regions under said conditions. Recently a non-invasive tongue based communication sys- tem has been developed, based around tongue-movement ear pressure (TMEP) signals [1]. The sensory unit is centered on a microphone positioned within the user’s external acoustic This research was supported by the UK Engineering and Physical Sciences Research Council (EPSRC), grant EP/F01869X Michael Mace is with the Department of Mechanical Engineer- ing at Imperial College London, London, UK, SW7 2AZ (email: m.mace11@imperial.ac.uk) Khondaker Abdullah-Al-Mamun and Shouyan Wang are with the Institute of Sound and Vibration Research (ISVR), University of Southampton, UK, SO17 1BJ (email: [kam1e08,sy.wang]@soton.ac.uk) Lalit Gupta is with the Department of Electrical and Computer Engineer- ing, Southern Illinois University, IL, USA, 62901 (email: lgupta@siu.edu) Ravi Vaidyanathan is with the Department of Mechanical Engineering at Imperial College London, London, UK, SW7 2AZ and the US Naval Postgraduate School, Monterey, CA, USA, 93943 (email: rxv@case.edu) meatus with user intention expressed through prescribed flicks of the tongue. These impulsive motions create unique low frequency (0 - 100 Hz) bio-acoustic pressure signals within the auditory cavity, allowing for inter-action dis- crimination and also discrimination from naturally occurring acoustic signals. Currently four actions have been defined and involve placement of the tip of the tongue at the base of the central incisor, left or right first molar and flicking the tongue up (bottom/left/right action) and placing the tip of the tongue against the top of the palate and flicking down (top action). This action set was chosen due to the tongue motions not normally occurring in daily activity, yet the actions themselves feel natural whilst executing, ensuring re- peatability. Previous work has shown inter-class classification results of various algorithms using four a-priori collected data-sets [1], discrimination between controlled and non- controlled movements based on the signal frequency content extracted using a wavelet packet transform [2] and initial real-time classification across three of the actions [3]. All this work has concentrated on mono-channel classification using individual classifiers. As an extension, an augmented bio-acoustic system based on a multi-channel ensemble classification framework is proposed. As opposed to obtaining data from a single micro- phone, a three channel system is implemented, consisting of a microphone placed within each ear and an acoustic gel sensor secured to the underside of the chin [4]. The acoustic gel sensor, although capturing a similar type of signal, provides additional information as the acoustic wave is propagating through a different facial region, thus providing different signal characteristics and therefore additional information. Fig. 1 gives an overview of the system (top-left), example waveforms associated with each channel from a single action (top-right) and a block representation of the ensemble pro- cess utilized (bottom). A multi-class system naturally allows for classifier rejection when there are conflicting channel outputs, preventing misclassification when there is ensemble disparities. This is vital when there is increased intra-class variance due to testing in non-controlled environments. In this paper the ensemble multi-channel framework is outlined, with its effectiveness for inter-class classification of TMEP signals highlighted through comparison to individual classi- fication baselines. II. METHODOLOGY Combining of multiple channels for classification naturally lends itself to be formulated within an ensemble classification framework. An ensemble classifier methodology can be 978-1-4244-4122-8/11/$26.00 ©2011 IEEE 1733 33rd Annual International Conference of the IEEE EMBS Boston, Massachusetts USA, August 30 - September 3, 2011