Knowledge‑driven large‑scale augmented Granger causality for classifying cannabis dependence

Functional magnetic resonance imaging in the resting state provides thousands of potential connections but only a few hundred time points, creating a curse of dimensionality for causal modeling and topology inference of brain networks. We tested whether large-scale Augmented Granger Causality (lsAGC)—which projects the data onto principal components and fits an autoregressive model—can detect cannabis-related network changes more accurately and efficiently than the state-of-the-art PCMCI causal discovery method or a non-causal correlation baseline. Forty young adults (20 regular cannabis users, 20 non-using controls) from the Addiction Connectome Preprocessed Initiative were analyzed with the Harvard–Oxford atlas (118 regions). To avoid circularity we ranked connections using an a priori weighting scheme derived from cannabinoid-1-receptor density and addiction neurocircuitry, then trained support-vector machines on feature subsets of 10–100 edges. Across multiple Wilcoxon signed-rank comparisons lsAGC outperformed both alternatives in the low-to-moderate feature range, achieving optimal performance with evidence-based feature selection (AUC 0.845, accuracy 0.775, F1 0.765; p < 0.001 versus PCMCI). PCMCI showed competitive performance only when larger feature sets were retained and required substantially longer computation time, whereas lsAGC finished efficiently while maintaining superior classification accuracy in clinically relevant feature ranges. The most discriminative edges linked limbic reward structures with frontal executive regions, mirroring previous cannabis neuroimaging findings. These results indicate that lsAGC offers a practical, interpretable, and computationally efficient causal biomarker for cannabis-use classification and is well suited to other high-dimensional connectome studies where scan duration is constrained.