An unsupervised plug and play BCI with consumer grade hardware

In this work we use the classic P300 speller where the user is presented a grid of characters. Groups of characters are repeatedly highlighted while the user is asked to count silently when the attended symbol is flashed. This elicits an increase in the potential difference recorded in the EEG, ~300 ms after a target stimulus is presented. By discriminating between the presence or absence of this so-called P300 waveform, it is possible to determine the target symbol. Typically, the decoder which performs the classification is trained in a supervised manner and requires a tedious and time consuming calibration session. In contrast, we use an unsupervised machine learning technique to train the classifier. This technique eliminates the need for a calibration session and allows the speller to be used instantly. The calibration is omitted by training the classifier with the data recorded during actual use and correcting previously spelled characters when more data is received. It has been shown that this method can compete with state of the art supervised methods on an auditory event-related potential paradigm [1] when high quality amplifiers are used [2]. We investigate the use of the Emotiv EPOC, which is a low-cost, 14 channel device, to build a plug and play BCI. The P300 speller application is implemented in the BCPy2000 framework. In our experiments 7 subjects were asked to spell 44 characters making use of the original row-column paradigm. For every character a sequence of stimuli, in which every row and column of the grid is intensified once in random order, is repeated for 15 iterations. Despite of the low signal to noise ratio of the Emotiv EPOC headset and the resulting decrease in accuracy for classifying P300 waveforms, reported in [3], a decent spelling accuracy is achieved making use of unsupervised machine learning. On average 81.8% (std. 11.74%) of the characters is correctly spelled after correction. Before correction 73% of the characters is correctly spelled. The majority of the originally misspelled and corrected characters are situated in the beginning of the sentence, where the data availble for training is still scarce. This scarcity can be solved making use of transfer learning to initialize the speller with basic information [4]. The exclusion of the calibration session makes the speller more attractive for use. Besides that, the results show that it is possible to build a speller with an affordable EEG recording system. We can conclude that we are one step closer to building a reliable, user friendly and affordable BCI system. References: [1] M. Schreuder, T. Rost and M. Tangermann. Listen, You are Writing! Speeding up Online Spelling with a Dynamic Auditory BCI. Frontiers in Neuroprosthetics, 5(112), 2011. [2] P. Kindermans, M. Schreuder, B. Schrauwen, K. Muller, M. Tangermann. True Zero-Training Brain-Computer Interfacing U+2013 an Online Study. Submitted [3] M. Duvinage, T. Castermans and T. Dutoit. A P300-based quantitative comparison between the Emotiv EPOC headset and a medical EEG device. Proc. 9th IASTED Int. Conf. on Biomedical Engineering, pages 1-8, 2012. [4] P. Kindermans, M. Tangermann, K. Muller, B. Schrauwen. Integrating Dynamic Stopping, Transfer Learning and Language Models in an Adaptive Zero-Training ERP Speller. Journal of neural engineering, in press

Jaar van publicatie:2014