Neural Decoding Requirements for a Take-home Brain Computer Interface David A. Friedenberg, Michael Schwemmer, Nicholas Skomrock, Per Sederberg, Jordyn Ting, Marcia Bockbrader, and Gaurav Sharma
Abstract Brain computer interfaces (BCIs) have had several successful laboratory demonstrations, raising hopes that a take-home system could improve the lives of patients in the future. However, challenges remain in translating BCI control of an assistive device in the lab into a robust take-home system. One challenge is designing neural decoders, algorithms that translate neural activity into control commands for a device, that meet BCI systems and extract requirements for neural decoding.
Translating laboratory demonstrations of BCI systems to home-use requires careful consideration of patient priorities and signficant technical challenges. To address the former, potential users were surveyed on several BCI characteristics. We examine two such surveys and extract requirements for the technical challenge of BCI decoding, the algorithms that translate brain activity into actions. In one survey, potential users ranked non-invasiveness, daily set-up time, independent operation, cost, number of functions provided, and response time as 2-)/ $'&)+%+4* %)+) *+ * [1]. In another, the number of functions, simplicity of setup, accuracy, electrode type, setup time, and speed were all ranked with a median importance of least 9 out of 10 [2]. The characteristics directly impacted by the decoding algorithm fall into four main categories: set-up time related to decoder training, number of functions provided, response time, and accuracy. The decoder-related setup time is primarily the time spent by the user calibrating the decoding algorithm to account for any day-to-day variability in the neural signals. It is typically performed on a pre-defined task where data is collected to train or update the decoding algorithms as the user performs the guided task. In [1] users expressed willingness to spend more time during initial training but want to minimize or even eliminate daily decoder set-up time. Overall, a total set-up time of 10-20 minutes would satisfy 65% of potential users, however that includes set-up time not related to decoding [2]. The number of functions enabled by a BCI varies based by device, but can be measured as the number of discrete movements (e.g. [3]) or the degrees of freedom for controlling *Research supported by Battelle and the Ohio State University. D.Friedenberg, M. Schwemmer, N. Skomrock, J.Ting and G. Sharma are with Battelle, Columbus, OH 43201 USA (email:
[email protected]). M. Bockbrader is with Center for Neuromodulation and Department of Physical Medicine and Rehabilitation, The Ohio State University, Columbus OH 43210 P. Sederberg is with Department of Psychology, University of Virginia, Charlottesville, VA 22904
an assistive device (e.g. [4]). The number of functions desired by potential BCI users has not been studied in the literature, but it can be presumed that more is better. The response time of a BCI system can be defined as the lag between the user intending to act and the BCI executing the action. For certain applications, there are other measurements of response time, for instance letters typed per minute in a BCI used for communication. In [2], 55% of respondents said a typing rate of 20-24 letters per minute would be acceptable while in EEG-based BCIs, users report a loss in sense of agency with response times as short as 750ms [5]. Ideally, a BCI system would always correctly interpret the ,*)5* %+%+,+ +. ## %- +#/$"$ *+"*%[2], 90% of users say an accuracy of 90% would be acceptable. The exact way accuracy is defined may differ depending on the function of the BCI [6] but a standard accuracy metric is the percentage of trials in which the user was successful (e.g. [7]). In conclusion, BCI decoding algorithms should be designed with user priorities in mind, namely minimizing decoder-related setup time, performing a reasonable number of functions, short response times, and high accuracy. However, optimizing all four simultaneously can be difficult as several of the priorities can conflict, for instance increasing the number of movements may decrease accuracy and one way to increase accuracy is increasing the training time [6] . REFERENCES [1] J. L. Collinger et al., 2 ,%+ &%#) &) + *** *+ -%&/% Brain-&$',+)%+)*+)' %#&)%!,)/4J. Rehabil. Res. Dev., vol. 50, no. 2, pp. 1451160, Apr. 2013. [2] J. E. Huggins, A. A. Moinuddin, A. E. Chiod&%)%2+ Would Brain-Computer Interface Users Want: Opinions and Priorities of &+%+ #*)* +' %#&)%!,)/4Arch. Phys. Med. Rehabil., vol. 96, no. 3, Supplement, p. S381S45.e5, Mar. 2015. [3] C. E. Bouton et al.2*+&) %&rtical control of functional movement in ,$%. +(,) '# 4Nature, vol. 533, no. 7602, pp. 2471250, May 2016. [4] J. E. Downey, L. Brane, R. A. Gaunt, E. C. Tyler-Kabara, M. L. Boninger, % # %)2&+&)&)+ #+ - +/%*,) %g neuroprosthetic-&%+)#&!+ %+)+ &%4Sci. Rep., vol. 7, no. 1, p. 16947, Dec. 2017. [5] -%* #,))%#%"2 *,# " Dominates the Sense of Agency for Brain- %+ &%*4PLOS ONE, vol. 10, no. 6, p. e0130019, Jun. 2015. [6] &$*/*&%%#)2%%#/* *&')&)$% evaluation for motor- $)/*4J. Neural Eng., vol. 10, no. 3, p. 031001, May 2013. [7] D. A. Friedenberg et al.2,)&')&*++ -enabled control of graded arm m,*#&%+)+ &% %')#/0,$%4Sci. Rep., vol. 7, no. 1, p. 8386, Aug. 2017.