Honors Program

University Honors

Date of Award


Thesis Professor(s)

Eric Sellers

Thesis Professor Department


Thesis Reader(s)

Paul Tudico


Incremental improvements are continuously being made to P300-Speller BCI paradigms. Accurate classification depends on a high signal-to-noise ratio (SNR) between the target and nontarget items. Fixed presentation rates produce a large flash-evoked response that persists throughout the recording epoch, which can potentially undermine the classification of P300-responses. By introducing a random interstimulus interval (ISI) to a previously improved P300-Speller paradigm (i.e., Checkerboard Paradigm; CBP) we expect to reduce the deleterious flash-evoked responses and increase the P300 classification SNR. Data were recorded from 32 EEG locations (right mastoid referenced) from 13 subjects using the CBP with two conditions. In the Random ISI (RI) condition, ISI varied between 0 ms and 187.5 ms and averaged 93.75 ms. In the Fixed ISI (SI) condition, ISI remained static at 93.75 ms. In both conditions, participants were instructed to spell out 72 characters using an 8x9 matrix of alphanumeric characters by silently counting each target flash. The first 36 characters served as ‘calibration’ data for a stepwise linear discriminant analysis (SWLDA; 0 - 800 ms poststimulus epochs). This SWLDA classifier was then used to provide online feedback for an additional 36 character selections. Absolute amplitude of target and nontarget responses were summed across the recording epoch for each subject and averaged between Pz and Cz (maximum). Target averages were then divided by nontarget averages to create a SNR measure and compared between RI and FI conditions. The RI manipulation produced a significantly (p = .04) larger SNR (M = 5.85) than the FI condition (M =4.07).Further analysis of the averaged waveforms revealed a significantly (p = .05) greater positive peak at Cz (253 ms peak latency) for the RI condition. Classification performance measures for RI and FI conditions were high for accuracy (84 and 85%, respectively; NS) and bitrate (21 and 23 bits/min, respectively; NS). Together these results suggest that while randomizing ISI can yield higher SNR, response classification is not affected. It is possible that SWLDA is a useful classification method, in general; however, these data suggest that it does not capitalize on the additional information gained from the increase in SNR. Alternative classification techniques that can take advantage of specific subcomponents of the response may be able to utilize this additional information to improve BCI speed and accuracy.

Document Type

Honors Thesis - Open Access

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 3.0 License.


Copyright by the authors.