BiN-STG-KAN: Integrating biologically-informed spatio-temporal graphs with Kolmogorov–Arnold networks for chronic low back pain decoding

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).

Chronic low back pain (CLBP) has been associated with altered large-scale brain networks, but subject-level discrimination from resting-state EEG remains difficult because of low signal-to-noise ratio, volume conduction, and inter-individual variability. We developed BiN-STG-KAN, a biologically informed spatio-temporal graph model with a Kolmogorov–Arnold network decision module, to characterize EEG connectivity patterns associated with CLBP case-control status. Orthogonalized amplitude envelope correlation (AEC-c) and debiased weighted phase lag index (dwPLI) were used to estimate connectivity, while Yeo-7 network masks constrained graph construction. The model was evaluated in 127 participants (72 CLBP and 55 healthy controls) using five-fold outer stratified subject-level cross-validation with three-fold inner hyperparameter optimization. BiN-STG-KAN achieved an accuracy of 84.4% and a ROC-AUC of 0.854. In exploratory CLBP-only analyses, held-out out-of-fold probabilities were not associated with pain intensity or pain duration after false-discovery-rate correction, although robust regression showed a modest positive association with pain intensity. Attribution analysis showed the largest aggregate contribution from γ -band streams, followed by α - and θ -band streams, while AEC − and dwPLI were the leading connectivity mechanisms; these patterns were consistent across outer folds and random initializations. These results identify EEG connectivity patterns associated with CLBP case-control status in this dataset but do not establish pain-severity prediction, disease specificity, external generalizability, or clinical utility. External and longitudinal validation with richer clinical phenotyping is required.