Dissertation Date: August 25, 2026
Functional near-infrared spectroscopy (fNIRS) measures the same underlying neurovascular response as functional magnetic resonance imaging (fMRI), but with inexpensive, compact, relatively motion-tolerant bedside instrumentation. This suits it to clinical questions requiring only coarse localization or detection of a regional response, such as presurgical mapping of eloquent cortex and detection of covert consciousness. Clinical questions concern a single patient, whereas most fNIRS methodology was developed at the group level, where averaging absorbs certain errors and variability. This dissertation tests whether whole-head, source-localized fNIRS can provide statistically valid, repeatable, and spatially accurate localization in individual subjects, the condition for its use as a bedside alternative to fMRI.
Aim 1 examined statistical validity of individual-level analyses. A systematic review of 234 fNIRS studies found that 78% applied frequency filtering, most commonly a 0.01–0.1 Hz band-pass, typically with tests assuming temporally independent samples. A factorial simulation crossing three data types, four filtering conditions, and six analysis methods showed this combination to be anticonservative for individual-level analysis: false positive rates for t-tests, ANOVA, and OLS general linear models reached 73% against a nominal 5%. Autoregressive prewhitening with iteratively reweighted least squares (AR-IRLS) restored nominal control on unfiltered data, with the best power of any valid method, but failed after low- or band-pass filtering; phase-randomized bootstrapping remained valid throughout, at a cost in power.
Aim 2 examined the precision of individual source localization and its robustness to probe placement. Five reconstruction algorithms were evaluated across 61 simulated probe placements over the 82 parcels detectable by a whole-head montage, and against ten-session test-retest recordings in four subjects. At exactly known placement, sLORETA, depth-weighted minimum-norm estimation, and dynamic statistical parametric mapping (dSPM) localized a parcel to within about 5 mm, clearly separated from LORETA and SFLEX. All algorithms degraded significantly under simulated probe rotation, median localization error rising from roughly 5 mm to 9 mm between mild and severe displacement. dSPM was the most robust in simulation and in test-retest recordings, where its 10.4 mm session-to-session centroid displacement bettered every other algorithm.
Aim 3 compared whole-head, source-localized fNIRS with same-day fMRI in 18 subjects, using a 102-channel montage, digitized optode positions, and each participant's anatomy. Median within-subject Dice coefficients were 0.55 for the motor task and 0.60 for the visual task, with individual values from 0.13 to 0.81. Median centroid displacements of the primary cluster were 6.2 mm (motor) and 6.3 mm (visual).
Together the aims show that valid single-subject inference is achievable with existing methods. Session-to-session precision is roughly one centimeter when probe placement varies, and about 5 mm when it is known exactly. Within-subject agreement with fMRI is about 6 mm when optode positions are digitized, comparable to the published 5–10 mm median center-of-mass displacement reported for motor and visual paradigms in clinical fMRI test-retest; spatial overlap (median Dice 0.55, finger tapping) likewise approaches published values for fMRI's own between-session overlap for the same task (0.574). The results support fNIRS use in individual subjects and its viability as a clinical tool, especially where the probe can be exactly registered to subject anatomy.
