Large‑scale Granger causality analysis of cannabis‑related network changes in the resting state with informed priors
A. Wismüller, A. Vosoughi, Akhil Kasturi
Medical Imaging April 1, 2026 DOI: 10.1117/12.3086350 (opens in new tab)
Study at a glance
AI-extracted from the abstract| Characteristics | Observational case-control classification study Peer reviewed |
|---|---|
| Sample size | 40 |
| Population | Young adults, half regular cannabis users and half non-users |
| Topics | Cannabis |
| Key findings | Large-scale Granger Causality (lsGC) detected cannabis-use-related brain network differences more accurately and efficiently than PCMCI or correlation, particularly with 20-50 selected connections (AUC 0.803, accuracy 0.770, F1 0.756). The most discriminative connections involved limbic reward and frontal executive regions, consistent with prior cannabis neuroimaging findings. |
Abstract
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 Granger Causality (lsGC)—which uses principal component analysis and autoregressive modeling—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. The Support-vector machines were then trained on subsets of 10 to 100 edges. The results of multiple Wilcoxon signed-rank tests showed that lsGC significantly outperformed both PCMCI and the correlation approach in the low-to moderate feature range (20-50 features), achieving optimal performance with evidence-based feature selection (AUC 0.803, accuracy 0.770, F1 0.756; p ⪅ 0.001). PCMCI exhibited lower performance (AUC 0.592) and required more computation time, while the correlation method was ineffective in all feature ranges (AUC 0.562). 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 lsGC is a practical, interpretable, and computationally efficient causal biomarker to classify cannabis use and is promising for other high-dimensional connectome studies with limited scan durations.
Comparable studies
Other observational and cohort studies on cannabis, most cited first.
| Study | Year | Design | Participants |
|---|---|---|---|
| High-potency cannabis and the risk of psychosis People with first episode of psychosis and healthy controls from the local population | 2009 | Case-control study | n = 454 |
| Ayahuasca-Assisted Therapy for Addiction: Results from a Preliminary Observational Study in Canada Rural aboriginal population | 2013 | Observational study | |
| Impaired cognitive performance in drug free users of recreational ecstasy (MDMA) Otherwise healthy young people | 2000 | Observational | |
| Continued cannabis use and risk of incidence and persistence of psychotic symptoms: 10 year follow-up cohort study Individuals from the general population in Germany aged 14-24 at baseline | 2011 | Observational cohort | n = 1,923 |
| Cannabis-induced psychosis and subsequent schizophrenia-spectrum disorders: follow-up study of 535 incident cases Patients treated for cannabis-induced psychotic symptoms in Denmark between 1994 and... | 2005 | Observational cohort | n = 535 |