Artifact filtering application to increase online parity in a communication BCI: progress toward use in daily-life

Abstract

A significant challenge in developing reliable Brain-Computer Interfaces (BCIs) is the presence of artifacts in the acquired brain signals. These artifacts may lead to erroneous interpretations, poor fitting of models, and subsequent reduced online performance. Furthermore, BCIs in a home or hospital setting are more susceptible to environmental noise. Artifact handling procedures aim to reduce signal interference by filtering, reconstructing, and/or eliminating unwanted signal contaminants. While straightforward conceptually and largely undisputed as essential, suitable artifact handling application in BCI systems remains unsettled and may reduce performance in some cases. A potential confound that remains unexplored in the majority of BCI studies using these procedures is the lack of parity with online usage (e.g., online parity). This manuscript compares classification performance between frequently used offline digital filtering, using the whole dataset, and an online digital filtering approach where the segmented data epochs that would be used during closed-loop control are filtered instead. In a sample of healthy adults (n = 30) enrolled in a BCI pilot study to integrate new communication interfaces, there were significant benefits to model performance when filtering with online parity. While online simulations indicated similar performance across conditions in this study, there appears to be no drawback to the approach with greater online parity.

Keywords: EEG; N200 and P300 potentials; artifact handling; brain computer interface; cBCI; online parity; signal filtering; signal processing.

Figure 1

RSVP calibration task. The RSVP calibration task prompts a user to search for a letter in an inquiry. This one-second prompt is followed by a fixation of 0.5 s, and then the 10-character inquiry. Each inquiry was presented at a rate of 5 Hz for a total time of two seconds per inquiry. The full iteration lasts 3.5 s. The user completed this process 110 times with a four second blank screen between inquiry iterations.

Figure 2

Online and conventional filter application. This study examines two filtering pipelines, each with several similar steps. However, the OF condition requires an additional epoching step before filtering. The top of figure illustrates this distinction. In the OF condition, raw data is first epoched into Inquiry Data before filtering, then further epoched into Trials, and finally passed into a Signal Model. In contrast, the CF condition filters the data immediately, epochs it into trials, and then passes it to a Signal Model for training.

Figure 3

Filter application on EEG signals. (A) Grand average of calibration data (n = 30) showing the averaged ERPs recorded from channels ‘Pz’, ‘Cz’, ‘Oz’, ‘P3’, ‘P4’, ‘O1’, ‘O2’. This panel shows that the paradigm successfully evoked a P300 response during the target condition that was classifiable for use in online typing. In the target conditions, the orange line represents the conventionally filtered data (CF) and the blue line the online filter (OF). In the non-target conditions, the red line represents CF and the green line OF. There appears no major distinction on group average for the ERP. (B) Topographic maps of target condition for CF (left) and OF (right). These demonstrate activity across channels and the impact of the different filtering conditions across channels. The OF produces similar results to CF with some reduction in early potentials (N1, P1) and slight changes in topography.

Figure 4

Filter band settings on classification performance. This figure demonstrates the impact of filter band settings on classification performance between filter application conditions. All OF results are shown in grey; CF results are shown in black. Performance was determined with the default PRK model and measured in terms of MCC. Standard error bars applied. Averages are reported at the bottom of each condition. The OF condition provided better or equal performance to the CF condition across all filters. The 1-20 Hz filter performed the best on this dataset across conditions.

Figure 5

Online simulation results. The above notched box plots demonstrate stimulated online classification performance between filter application conditions across participants. The means are plotted using a green triangle, outliers are denoted with circles above/below the whiskers. In the top plots, MCC (left) and BA (right) are plotted. The models performed similarly with OF having a slight advantage on average.

Figure 6

Online simulation confusion matrices. This figure presents the averaged confusion matrices for the online simulations in the OF (left) and CF (right) conditions. While there were no statistical differences in the primary performance measures (BA/MCC), slight differences in model performance can be seen on average. The OF model predicted more of the positive class on average, where CF favored the negative class.