Comparing different pre-processing routines for infant fNIRS data

Functional Near Infrared Spectroscopy (fNIRS) is an important neuroimaging technique in cognitive developmental neuroscience. Nevertheless, there is no general consensus yet about best pre-processing practices. This issue is highly relevant, especially since the development and variability of the infant hemodynamic response (HRF) is not fully known. Systematic comparisons between analysis methods are thus necessary. We investigated the performance of five different pipelines, selected on the basis of a systematic search of the infant NIRS literature, in two experiments. In Experiment 1, we used synthetic data to compare the recovered HRFs with the true HRF and to assess the robustness of each method against increasing levels of noise. In Experiment 2, we analyzed experimental data from a published study, which assessed the neural correlates of artificial grammar processing in newborns. We found that with motion artifact correction (as opposed to rejection) a larger number of trials were retained, but HRF amplitude was often strongly reduced. By contrast, artifact rejection resulted in a high exclusion rate but preserved adequately the characteristics of the HRF. We also found that the performance of all pipelines declined as the noise increased, but significantly less so than if no pre-processing was applied. Finally, we found no difference between running the pre-processing on optical density or concentration change data. These results suggest that pre-processing should thus be optimized as a function of the specific quality issues a give dataset exhibits.