Investigation of the sensitivity-specificity of canonical- and deconvolution-based linear models in evoked functional near-infrared spectroscopy.

Functional near-infrared spectroscopy (fNIRS) is a noninvasive brain imaging technique to measure evoked changes in cerebral blood oxygenation. In many evoked-task studies, the analysis of fNIRS experiments is based on a temporal linear regression model, which includes block-averaging, deconvolution, and canonical analysis models. The statistical parameters of this model are then spatially mapped across fNIRS measurement channels to infer brain activity. The trade-offs in sensitivity and specificity of using variations of canonical or deconvolution/block-average models are unclear. We quantitatively investigate how the choice of basis set for the regression linear model affects the sensitivity and specificity of fNIRS analysis in the presence of variability or systematic bias in underlying evoked response. For statistical parametric mapping of amplitude-based hypotheses, we conclude that these models are fairly insensitive to the parameters of the regression basis for task durations > 10    s and we report the highest sensitivity-specificity results using a low degree-of-freedom canonical model under these conditions. For shorter duration task ( < 10    s ), the signal-to-noise ratio of the data is also important in this decision and we find that deconvolution or block-averaging models outperform the canonical models at high signal-to-noise ratio but not at lower levels.