QUANTIFYING AND COMPENSATING FOR P300 VARIATIONS IN AMYOTROPHIC LATERAL SCLEROSIS

P300-based brain–computer interface (BCI) systems enable people with neuromuscular disabilities, including amyotrophic lateral sclerosis (ALS), to communicate and to control their environments using brain activity. However, BCI systems have not yet fulfilled their promise as reliable communication systems for all who need them. Despite continued work on improving BCIs for end users, people with ALS can experience both reduced performance overall compared to neurotypical users and significant day to day variations in BCI performance and event-related potential (ERP) characteristics. The P300 response, which the P300 speller is based on, is also known to exhibit trial-to-trial latency variability. The importance of latency jitter is established in cognitive studies, and its relevance to BCIs is of growing interest. Increased latency jitter is associated with decreased BCI performance, and preliminary comparisons indicated that jitter may be increased in ALS, similar to increased jitter found in a variety of neurological conditions. Therefore, we quantify latency jitter and its correlates in people with ALS, longitudinally investigate within-session variability in event-related potentials (ERPs), session-average ERPs, and their relationships, and develop and evaluate a correction method to compensate for latency jitter in BCI use. To this end, we use longitudinal EEG data collected from 6 participants with ALS, and, when applicable, from neurotypical control participants, using a P300 BCI. Data recorded in each session had session-average ERP amplitudes and latencies extracted. Stepwise linear discriminant analysis was used both to evaluate BCI performance and to support the use of classifier- based latency estimation (CBLE) to estimate whole-epoch latency shifts for single trials in all aims. To quantify latency jitter and its correlates in people with ALS, latency jitter was calculated with CBLE. Then, ERP components and latency jitter were compared between participants with ALS and neurotypical control participants using Wilcoxon rank-sum tests. Correlations between latency jitter and each of the clinical measures, ERP features, and performance measures were investigated using Spearman and repeated measures correlations. We found that latency jitter was significantly increased in participants with ALS and significantly negatively correlated with BCI performance in both ALS and control participants. We also found significant correlations between ERP amplitudes and latency jitter in neurotypical participants and reduced ERP amplitudes in participants with ALS. However, there was no significant correlation between latency jitter and clinical measures. Based on these results, we proposed a data augmentation and jitter correction (A/C) scheme with parameters determined individually using latency shifts calculated with CBLE. Performance metrics including character selection accuracy and binary accuracy, precision, recall, and F-score were calculated using both the proposed classification scheme and a reference classifier that did not implement data augmentation or correction. Performance was compared between the two classification methods using paired t-tests and investigated longitudinally using correlation analyses. Correlations between performance improvements and clinical measures were also investigated. The proposed classification scheme significantly improved character selection accuracy, required for usability, as well as recall and F-scores. However, precision was reduced, and binary accuracy was not significantly affected. Overall, BCI performance deteriorated over time with both classification methods, and latency jitter calculated with CBLE increased over time. Improvements in selection accuracies using the proposed A/C approach were greater for participants with more significant physical impairments. Also following the results from the first aim, we extracted single-trial N100, P200, N200, and P300 amplitudes and latencies in each session using Woody-type filters on spatially filtered data. That is, spatial principal component analysis was conducted on the responses to stimuli containing the intended characters in each session, and appropriate spatial factors were selected from the results of this analysis. Then, session-average time series for these spatial factors were used as templates. The cross-covariance of the templates with the single-trial time series were calculated. The maximum value of these cross-covariances and the latency shifts to achieve this maximum were then used as single-trial amplitudes and latencies. Within-session variability in N100, P200, N200, and P300 latencies were compared between participants with ALS and neurotypical participants using Wilcoxon rank-sum tests, and P200, N200, and P300 jitter were all found to be increased in ALS. In addition, linear models were used to investigate which ERP feature latencies contributed to the shifts detected with CBLE, determining that single-trial N100, P200, N200, and P300 latencies were all significant contributors in data recorded from neurotypical participants. However, the relationships between ERP feature latencies and CBLE were disrupted in ALS, with single-trial N100 latencies no longer a significant contributor to latency shifts calculated with CBLE and reduced but still significant contributions from single-trial P200, N200, and P300 latencies. There were, however, some contributions to jitter from single-trial ERP amplitudes, with increased latency shifts detected with both CBLE and Woody filters on trials with reduced ERP amplitudes. Considering these results, we conclude that CBLE reflects both latency jitter and other factors which affect BCI performance. Despite the increase in latency jitter calculated with CBLE over time in participants with ALS, there was not a significant increase in N100, P200, N200, or P300 jitter calculated with Woody filters over time. Overall, the research presented in this dissertation advances knowledge on latency variability in the use of P300 BCIs, both for neurotypical participants and for people with ALS. The importance of latency jitter in P300 BCIs is elucidated, both whole-epoch jitter calculated with CBLE and latency variation in specific ERP features are shown to be increased in people with ALS, a theoretical limitation of CBLE is investigated, and a compensation strategy is proposed to address increased latency jitter in people with ALS using P300 BCIs