Correction-stage information leakage and target-input asymmetry can distort comparisons of fNIRS short-separation channel correction strategies

We examined how evaluation design affects comparisons of functional near-infrared spectroscopy (fNIRS) short-separation correction strategies.

Nearest-SC, All-SC, SC-PCA and an adapted tCCA-GLM were compared in a multimodal resting-state dataset (n = 14), using semi-nested and correction-stage fully nested cross-validation, boundary purging, symmetric leave-one-modality-out (LOMO) evaluation, circular-shift nulls, injected hemodynamic response function (HRF) recovery and a second cohort.

In an initial joint-target diagnostic analysis, the adapted tCCA-GLM's absolute out-of-fold R² increased by 42.7 percentage points from semi-nested to purged fully nested evaluation. In a separate fixed-seed purged analysis, its null-adjusted ΔR² was -1.1%. Symmetric LOMO, which evaluates reduced tCCA models, yielded no statistically detected method differences (all paired p ≥ 0.3685), but did not establish equivalence or separately quantify leakage and input-target overlap. Motion-targeted external validation provided partial support for an evaluation-tier × method interaction (Freedman-Lane p < 0.005). SC-PCA had lower mean known-truth normalized root mean square error (nRMSE = 0.037) than the adapted tCCA-GLM (0.121); within-metric Holm-adjusted comparisons were significant for this contrast and for amplitude-bias comparisons with all three SC-based strategies, but not for slopes. Sensitivity analyses included training-fold PCA selection, LOMO boundary/segment-filtering checks, lagged evaluation and longer test blocks.

Benchmarks should fit learned correction stages within training data, disclose correction-evaluation overlap and assess signal preservation separately. Findings are specific to the evaluated implementations, outcomes and datasets; the sensitivity analyses do not establish complete removal of preprocessing-related dependence.