Detecting cannabis use with fMRI by applying large‑scale kernelized Granger causality and evidence‑based selection

Functional magnetic resonance imaging (fMRI) in the resting state often involves analyzing thousands of potential connections with only a few hundred time points, presenting a significant challenge in causal modeling and topology inference of brain networks. This study evaluates whether large-scale Kernelized Granger Causality (lsKGC)—which uses kernel methods with principal component analysis in feature space and autoregressive modeling to capture nonlinear causal dependencies—can more accurately and efficiently detect network changes associated with cannabis use compared to the PCMCI causal discovery method and a non-causal correlation baseline. The study included forty young adults, half of whom were regular cannabis users and half non-users, analyzed using the Harvard–Oxford atlas that covers 118 regions of the brain. To minimize bias, connections were ranked based on a predefined scheme related to cannabinoid-1-receptor density and known addiction neurocircuitry. Support vector machines were then trained on subsets of 10 to 100 edges. The results of multiple Wilcoxon signed-rank tests showed that lsKGC significantly outperformed both PCMCI and the correlation approach across most feature ranges, achieving optimal performance with evidence-based feature selection (AUC 0.90, accuracy 0.82, F1 0.80; p ⪅ 0.001). PCMCI exhibited substantially lower performance (AUC 0.59, accuracy 0.62, F1 0.61) and required significantly more computation time, while the correlation method was ineffective across all feature ranges (AUC 0.56, accuracy 0.53, F1 0.52). The most discriminative connections involved limbic reward structures and frontal executive regions, which aligns with previous findings in cannabis neuroimaging studies. These findings suggest that lsKGC is a practical, interpretable, and computationally efficient causal biomarker for classifying cannabis use and shows promise for other high-dimensional connectome studies with limited scan durations.