Non-linear functional topology for fNIRS-BCI: A DTW framework with explainable feature fusion and computational efficiency optimization

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Accurate motor intent decoding from functional near-infrared spectroscopy (fNIRS) signals is critical for deployable brain-computer interfaces (BCIs) in neurorehabilitation. Conventional feature extraction relies on linear functional connectivity metrics, like Pearson correlation, which are limited in their ability to accommodate the non-linear temporal variability of the hemodynamic response, constraining single-trial classification reliability. This paper presents a hybrid framework leveraging Dynamic Time Warping (DTW) to construct trial-level non-linear functional adjacency matrices, fused with localized hemodynamic magnitude features via an early-fusion architecture. The framework incorporates a leakage-free machine learning pipeline combining Minimum Redundancy Maximum Relevance (mRMR) feature selection—applied strictly within cross-validation folds—with Random Forest classification. Validation across two independent public datasets (52 subjects) performing coarse and fine-grained motor tasks demonstrated that the hybrid framework achieves a statistically significant improvement over linear Pearson baselines (p<0.0001, Cohen’s d=0.912, large effect). Furthermore, mRMR candidate feature analysis identified non-linear temporal couplings between motor cortex channels as the most consistently discriminative features relative to localized signal amplitude. Study 2 (Dataset B, N=22) is reported as a preliminary pilot exploration due to extreme data scarcity (3 trials per class per subject, achieved statistical power=7.0%). A two-stage computational optimization strategy—combining polyphase resampling to 2 Hz with selective topology computation restricted to the most discriminative channel pairs— reduced mean computational latency to 0.044 seconds (95th percentile: 0.051 s), demonstrating computational computational feasibility for future real-time BCI deployment. These results demonstrate that non-linear temporal topology constitutes a physiologically grounded, computationally efficient, and interpretable feature engineering solution for fNIRS-based assistive technologies.