Full Model Optimisation of the Processing Pipeline in Functional Near-Infrared Spectroscopy

Robert J. Ward, Gustavo Rodriguez-Gomez, Felipe Orihuela-Espina

Abstract: The analysis of functional Near-Infrared Spectroscopy (fNIRS) data relies on complex processing pipelines composed of multiple functions and parameters, yet the design of such pipelines remains ad hoc and heterogeneous across the field. This lack of standardisation undermines reproducibility and hinders objective comparison of findings. Here, we introduce a general optimisation framework for full model selection (FMS) of fNIRS pipelines, formulated as a hybrid discrete–dense optimisation problem. Our approach leverages gradient-based methods to jointly optimise function choices and parameter values, enabling systematic exploration of the mixed search space without reducing it to purely discrete approximations. To alleviate combinatorial explosion in the discrete search space, we introduce a graph pruning step that reduces the number of valid node connections without excluding admissible solutions. We validate the framework using synthetic and proof-of-concept NIRS datasets, demonstrating its ability to recover plausible pipelines and its flexibility with respect to user-defined criteria such as accuracy, physiological plausibility, or reproducibility. While illustrated in fNIRS, the framework is broadly applicable to neuroimaging and other domains where pipelines integrate categorical and continuous components. By shifting pipeline design from artisanal practice to principled optimisation, this work contributes to the development of reproducible and standardised neuroimaging workflows.

External IDs:doi:10.1007/s12021-026-09772-7