Abstract
Individual differences in brain anatomy are frequently used to predict cognitive function, yet the standard practice of accounting for intracranial volume (ICV) through proportional correction remains under-scrutinized in predictive modeling. This study evaluates how ICV correction impacts sex-independent and sex-specific models of cognitive ability across three anatomical properties: surface area, gray matter volume, and cortical thickness. Using data from the Human Connectome Project (HCP) and the Adolescent Brain Cognitive Development (ABCD) study, we demonstrate that ICV correction differentially biases results. While it improves predictive accuracy for cortical thickness, it reduces accuracy for surface area and gray matter volume. Furthermore, ICV correction significantly alters conclusions regarding the generalizability of models across sexes and age groups, often introducing artificial sex differences where none previously existed.
Introduction
In neuroimaging, the substantial variation in head size across individuals, sexes, and age groups is typically treated as a confounding variable. To isolate regional anatomical differences, researchers often employ "proportional correction," dividing regional measurements by the total intracranial volume (ICV). While this normalization is standard in morphometric analyses, its application in brain-behavior prediction is often performed without considering whether ICV itself contains biologically relevant signal. This paper investigates whether ICV correction aids or hinders the prediction of human cognitive abilities and how it influences the perceived similarity of neural architectures across different demographic subgroups.
Methodology
The researchers utilized a machine-learning framework to assess the impact of ICV correction on two large-scale datasets:
- Human Connectome Project (HCP): 1,013 healthy young adults (ages 22–37).
- Adolescent Brain Cognitive Development (ABCD) Study: 1,823 typically developing children (ages 9–10).
Cognitive performance was quantified using the NIH Toolbox Cognition Battery, focusing on crystallized and fluid intelligence. The predictive models utilized linear ridge regression to map three anatomical features—cortical surface area, gray matter volume, and cortical thickness—to cognitive scores. Models were trained using both "raw" (uncorrected) data and "proportional-corrected" data (Feature / ICV).
To ensure the interpretability of the models, the authors applied the Haufe transformation to model weights. This step ensures that the importance assigned to a specific brain region reflects its actual covariance with the behavior rather than mathematical artifacts. Finally, the study tested model translation, training models on one subgroup (e.g., males or adults) and testing them on another (e.g., females or children) to evaluate generalizability.
Key Results
The Impact on Predictive Accuracy
The study found that ICV is not merely noise; it correlates significantly with cognitive ability ($r \approx 0.16$ to $0.24$). Consequently, the effect of correction depends heavily on the anatomical feature being measured:
- Surface Area and Volume: In the adult HCP dataset, predictive accuracy ($r$) for surface area dropped from 0.190 (uncorrected) to 0.142 (corrected). Because these features are strongly tied to global brain size, correction removes behaviorally relevant information.
- Cortical Thickness: Conversely, accuracy for cortical thickness rose from 0.119 (uncorrected) to 0.174 (corrected). Since raw thickness is only weakly related to ICV, the division process effectively "injects" the behaviorally relevant ICV signal into the thickness measures, indirectly boosting the model's predictive power.
Generalizability and Sex Differences
ICV correction fundamentally changes how models generalize across populations. Models based on uncorrected surface area were significantly more effective at predicting scores across different age groups (e.g., using an adult-trained model on children) than their corrected counterparts.
Most strikingly, ICV correction appears to create a "generalizability trap" regarding sex. Without correction, male and female models for cognitive ability relied on highly similar brain features (correlation $r \approx 0.80$). However, after ICV correction, this overlap plummeted to $r \approx 0.24$. This suggests that many reported sex differences in neuroanatomical predictive features may be artifacts of normalization strategies rather than true biological divergence.
Developmental Divergence
The relationship between ICV and brain anatomy shifts during development. In adults, the association cortices—responsible for high-level cognition—are more strongly tied to ICV. In children, this relationship is more uniform across the cortex. This shift implies that the bias introduced by ICV correction is not static and varies depending on the developmental stage of the cohort.
Discussion and Conclusion
These findings serve as a cautionary tale for neuroimaging researchers. Proportional ICV correction is not a neutral preprocessing step; it can introduce bias where none existed and remove valid biological signals. The choice to correct for ICV should be driven by the specific anatomical feature and the population under study.
While this study focused on proportional correction, future research is needed to determine if alternative methods—such as residual-based correction or covariate regression—can mitigate these biases without the risk of data leakage. Ultimately, the study highlights that in the quest to normalize brain data, we must be careful not to discard the very individual differences we aim to understand.
Summary updated