Robust and automated motion correction for real infant fNIRS data

Although many approaches have been proposed, removing motion artifacts from developmental fNIRS data remains a difficult challenge. In particular, the lack of consistency in motion correction approaches across experimental reports suggests that the field has not yet identified an algorithm that consistently removes the majority of motion contamination while retaining hemodynamic responses, regardless of the idiosyncrasies of particular datasets. Some existing approaches remove the same fraction of variance from each participant’s data; others use participant data to set filtering parameters in ways that result in more stringent thresholds for low-motion participants than high-motion participants. Both types of approach risk leaving artifacts in data from participants with the most motion, while removing signal from participants with the least motion. In contrast, the procedure proposed here identifies and filters motion artifacts on the basis of a fixed, physiologically-justified threshold, so that amount of variance removed is closely associated with the prevalence of motion in each participant’s data. Across multiple contrasts from real experimental datasets, this procedure effectively removes motion artifacts while retaining the hemodynamic response signal, allowing the detection of differential responses to conditions, and recovering canonical hemodynamic response functions for both oxygenated and deoxygenated timecourses, indicated by robust negative correlations between the two hemoglobin types. This motion correction procedure would be appropriate to preregister as a planned component of the preprocessing stream in future fNIRS research.