A comparison of state-of-the-art signal classification algorithms in a locked-in end-user with a tactile P300 based BCI in long-term home-use

Objective: Brain-computer interfaces (BCIs) constitute a last resort of communication for people with severe motor impairments. To date, several potential improvements of BCI-related hardware and software components have been investigated, for instance, classification algorithms for brain signals such as electroencephalography (EEG). These efforts focused on laboratory work with healthy participants. However, this is not representative for real-life conditions of potential motor-impaired end-users with diseases like amyotrophic lateral sclerosis (ALS). Our study compares state-of-the-art classification approaches on a real-life dataset of an end-user with ALS and a tactile P300-based BCI, which was obtained in long-term home-use.Approach: We applied state-of-the-art classification approaches, including Shrinkage Linear Discriminant Analysis (ShrinkLDA) and Riemannian Geometry Classifiers (RGC). Moreover, we tested effects of different calibration scenarios (i.e., amount of training data, time between calibration) and spatial filtering (using xDawn). These variations were compared to the classification results of a commonly used Stepwise Linear Discriminant Analysis (SWLDA). As a criterion for sufficient performance, we defined an accuracy of >70% (based on prior work). Main results: SWLDA could be outperformed, but without spatial filtering via xDawn, no classification algorithm was able to allow for sufficient independent home-use (with accuracies ranging between 33% to 66%; SWLDA: 59%). The addition of spatial filtering via xDawn notably improved classification results for all classifiers, with the best performance achieved by RGC_xDAWN and ShrinkLDA_xDAWN (both 72% accuracy; SWLDA_xDAWN: 70%).Significance: This is the first study comparing classifiers on a large set of tactile P300-based BCI data obtained from an end-user in his own home. Of the state-of-the-art classifiers (established with healthy participants), only two reached slightly beyond a performance accuracy of >70%. Future work can build on our findings (e.g., implement spatial filtering) and further investigate the consequences of real-life conditions for potential end-users (e.g., disease characteristics, hardware and software optimization, environmental factors).