Abstract
Group analysis of brain magnetic resonance imaging (MRI) metrics frequently employs generalized additive models (GAM) to remove contributions of confounding factors before identifying cohort-specific characteristics. However, considering confounding factors before group analysis may remove information relevant for distinguishing specific cohorts. This study introduces Joi-GAM-Class, a machine-learning model that identifies neuromorphometric patterns by simultaneously training a GAM and a generic classifier. Using a baseline dataset from the National Consortium on Alcohol and NeuroDevelopment in Adolescence (NCANDA, N = 705), the joint model identified a unique pattern of eight brain regions in adolescents who misuse alcohol. This approach proved more accurate and impartial to confounding factors than traditional sequential methods, aligning with established alcohol literature.
Introduction
Studying the adolescent brain (ages 12–21) presents a unique challenge: the brain is a moving target. Structural changes—such as gray matter thinning and increased white matter organization—occur rapidly. When investigating the effects of alcohol misuse, researchers must distinguish between alcohol-induced changes and natural developmental variations or socioeconomic factors.
Traditional statistical approaches use a sequential two-step process: first, a Generalized Additive Model (GAM) "cleans" the data of confounds like age; second, researchers look for patterns in the residuals. This paper argues that this "pre-cleaning" can inadvertently strip away subtle signals necessary for accurate classification.
Methodology: The Joi-GAM-Class Model
The researchers proposed a joint optimization framework, Joi-GAM-Class (Joint GAM Classification). Unlike sequential models, this approach embeds the GAM directly into the classification's energy function.
Key features of the model include:
- Joint Optimization: Simultaneously determines parameters for confound removal and weights for a sparse logistic classifier, ensuring harmonization is informed by the classification goal.
- Sparsity Constraints: Forces the model to identify a small, interpretable subset of the most informative brain regions.
- Multimodal Integration: Analyzes 144 metrics across macrostructural MRI (volumes) and microstructural Diffusion Tensor Imaging (DTI).
Results
The study utilized the NCANDA baseline dataset, comparing 671 "minimal drinkers" to 34 "regular drinkers."
Classification Performance
Joi-GAM-Class achieved a normalized accuracy of 75.9%, significantly outperforming sequential GAM approaches. Crucially, the model passed an "age-test"—it could not distinguish between older and younger non-drinkers, proving it was truly age-blind and not simply using age as a proxy for drinking status.
The Eight-Region Signature
The model identified a stable neuromorphometric pattern involving eight specific regions: the lateral ventricles, mid-posterior and central corpus callosum, centrum semiovale, anterior corona radiata, posterior thalamic radiation, fornix, and cingulate gyrus.
Regular drinkers exhibited smaller volumes in the mid-posterior corpus callosum and centrum semiovale, alongside higher axial diffusivity in the fornix. These findings reinforce existing literature regarding alcohol's impact on white matter integrity during adolescence.
Discussion and Conclusion
The core insight of this research is that data harmonization should not be task-agnostic. By allowing the classification task to influence how confounds are removed, Joi-GAM-Class preserved critical group differences that sequential methods discarded as noise.
While the small sample size of the drinking cohort (N=34) and the cross-sectional nature of the baseline data are limitations, the study provides a robust framework for future longitudinal analysis. The findings suggest that joint optimization is a superior method for uncovering the subtle structural signatures of substance use in the developing brain.
Summary updated