Optimizing short-channel regression in fNIRS: an empirical evaluation in a naturalistic multimodal paradigm

Functional Near-Infrared Spectroscopy (fNIRS) is increasingly favored for its portability and suitability for ecological paradigms, yet methodological standardization remains a challenge, especially regarding the optimal use of short-separation channels (SC) to remove systemic physiological noise. While SC signals are widely recognized for isolating extracerebral hemodynamics such as cardiac and respiratory artifacts, no consensus exists on how best to incorporate them into analysis pipelines, limiting comparability across studies. Here, we systematically investigate six SC regression strategies within a Generalized Linear Model (GLM) framework, using fNIRS data acquired from temporal and occipital cortices of 16 healthy adults during naturalistic auditory, visual, and audiovisual passive stimulation. These strategies varied in the number of SCs used, anatomical specificity, and the application of dimensionality reduction. Our results show that for oxygenated hemoglobin signals, processing data without SC regression yields uninterpretable outcomes, with widespread negative beta estimates and no detectable task-related activation, underscoring the crucial role of SC correction. Among correction methods, approaches pooling all SC signals without anatomical constraints consistently outperformed spatially constrained techniques mimicking limited SC availability. Furthermore, applying orthogonalization to the full set of SCs provided additional improvements by efficiently capturing shared systemic variance and reducing redundancy, enhancing detection of modality-specific cortical activations and condition contrasts. Finally, although the deoxygenated hemoglobin signal is inherently less sensitive to systemic noise, it benefited from the same SC regression strategies, supporting their applicability across chromophores. Overall, our findings highlight the essential role of SC regression in recovering physiologically meaningful signals in fNIRS and recommend including all available SC channels within the GLM, coupled with orthogonalization techniques, as a robust and generalizable best practice for denoising across diverse experimental designs and hardware configurations.