Adaptive Brain-computer Interfacing through Error-related Potential Detection

This thesis investigates possible improved brain-computer interface (BCI) performance by using the error-related potential (ErrP) to guide adaptation in two forms: first by gradually adapting the algorithms that differentiate brain activity, and second by automatically correcting mistakes when observed. Five studies were conducted, each investigating a unique aspect of incorporating ErrPs into BCIs. In study one, gradual adaptation of linear classifiers differentiating left- and right-hand motor imagery was explored in computer simulation when labels were derived from a realistic, imperfect ErrP detector. Learning from all trials was compared to reactively learning from errors, and the former more consistently improved accuracies across a wider range of learning rates. In study two, a realistic error detector was biased to several false positive rates and the risk associated with ErrP-based adaptation of several motor imagery BCIs was evaluated in light of those biases. Reweighting of a classifier ensemble had lower risk than incremental adaptation of the ensemble's base classifiers, and a unique risk profile across error detector false positive rates for each adaptation method was discovered. In study three, automatic correction of a visual P300 speller was employed using an error detector that combined the ErrP and spelling confidence. Errors were detected with an average sensitivity of 86.67% and specificity of 96.59% (n=11), and automatic correction increased selection accuracy by 13.67 percentage points. In study four, ErrP-based error detection was used for gradual adaptation of visual P300 classifiers both when the stimulus timing pattern remained constant, and when stimulation speed doubled. By retraining when stimulation speed doubled using ErrP-based labels subject to a confidence score, online spelling efficiency increased from 1.77 to 2.63 symbols per minute (n=11). No significant improvement was found for this partially supervised adaptation when stimulation speed remained constant. In study five, ErrP detection was investigated in an auditory P300 BCI. ErrPs were detected with an area under the curve of 0.946 and simulated auto-correction improved ITR by 2.3 bits/min (n=9). These studies demonstrate promise for incorporating ErrPs into BCIs, particularly auditory P300 BCIs, provided false positives are minimized and BCI confidence is considered.Ph.D.2017-01-08 00:00:0