Evaluating Structural MRI and Alcohol-Use Covariates in Machine Learning Models of Heavy Cannabis Use: A Longitudinal Pilot Study

Machine learning is increasingly used to identify neuroimaging patterns associated with substance use, but small public datasets make it difficult to determine whether a model is learning brain-based signals or simpler non-imaging variables. This study evaluated whether T1-weighted structural magnetic resonance imaging (MRI) features improved the classification of heavy cannabis users versus healthy controls beyond demographic and alcohol-use covariates. T1-weighted MRI is an anatomical MRI scan type commonly used to visualize brain structure. In this paper, covariates refer to non-MRI variables that may help explain group differences, including age, gender, and baseline Alcohol Use Disorders Identification Test (AUDIT) score. A higher AUDIT score indicates greater alcohol-use risk or severity. The analysis used OpenNeuro ds000174, a public longitudinal dataset containing baseline and three-year follow-up MRI scans from 20 heavy cannabis users and 22 healthy controls, along with Cannabis Use Disorder Identification Test (CUDIT) scores, where higher CUDIT scores indicate greater cannabis-use problems. Baseline MRI and longitudinal MRI-change features were generated through intensity normalization, resizing, voxel flattening, and principal component analysis (PCA). Logistic regression was evaluated under stratified five-fold cross-validation. The covariate-only model achieved the strongest performance, with balanced accuracy of 0.620 and receiver operating characteristic area under the curve (ROC-AUC) of 0.608. Baseline MRI, longitudinal MRI change, combined MRI features, and MRI-plus-covariate features did not improve performance. These findings suggest that, in this limited public dataset, simple voxel-based structural MRI representations did not provide stable classification value beyond demographic and alcohol-use covariates.