MetaIDNet: A Fast Brain-Computer Interface for Color Vision Assessment

Our laboratory has recently demonstrated objective, quantitative, and precise brain-computer interface (BCI)-based color vision (CV) assessment. During assessment, participants look at a flashing stimulus that alternates between two light sources. Light sources that appear to be different colors elicit larger steady-state visual evoked potentials (SSVEPs); those that appear to be similar colors elicit smaller SSVEPs. Thus, the stimulus-evoked responses depend on an individual's CV. BCI-based CV assessments evaluate CV by observing SSVEP sizes elicited over a color space; their accuracy depends on the sampling strategy and noise in the individual observations. Existing systems uniformly sample the color space and average multiple SSVEPs observed from the same location to reduce noise. Effectively, these systems collect additional observations to improve assessment accuracy, but at the expense of assessment speed. We present MetaIDNet, a closed-loop BCI that uses Gaussian process regression (GPR) and a neural network to improve the speed-accuracy tradeoff of BCI-based CV assessment. We perform offline and online experiments with ten and eight participants, respectively, to evaluate the speed and accuracy of MetaIDNet and compare it with two existing methods. Our offline experiments show that MetaIDNet improves CV assessment speed by 70% (while maintaining accuracy), and our online experiments show that MetaIDNet improves CV assessment accuracy by 40% (while maintaining the same speed). The results confirm that MetaIDNet significantly improves the speed and accuracy of BCIs for CV assessment.