✨ TL;DR
This paper proposes a feature whitening approach that decorrelates anatomically grouped brain regions to improve the interpretability of linear neuroimaging models while maintaining predictive performance. The method helps disentangle overlapping information across correlated brain measures, making model weights more clinically meaningful for identifying biomarkers in psychiatric disorders.
Linear models are commonly used in computational neuroimaging to identify brain biomarkers associated with pathologies, but interpreting the learned weights is challenging because they often fail to yield clinically meaningful insights. This difficulty stems from strong inherent correlations between brain regions—particularly homologous structures in left and right hemispheres—which causes linear weights to reflect shared contributions across regions rather than region-specific effects. Consequently, standard linear models obscure the true neurobiological mechanisms underlying predictions.
The authors introduce a whitening approach that leverages prior neuroanatomical knowledge to decorrelate groups of brain regions with known shared variance. The method specifically targets anatomically informed pairs (such as homologous hemispheric structures) to disentangle overlapping information. They also propose a regularized variant that allows controlled tuning of the decorrelation degree. Critically, unlike dimensionality reduction techniques such as PCA or ICA, this approach retains the full input signal while decorrelating features, making it specifically designed for feature interpretation rather than feature selection.
What the paper shows.
The method was evaluated on region-of-interest features in two psychiatric classification tasks: distinguishing individuals with bipolar disorder from healthy controls and distinguishing individuals with schizophrenia from healthy controls. The findings demonstrate that whitening improves the interpretability of model weights while preserving predictive performance, providing a robust framework for linking linear model outputs to neurobiological mechanisms.
The paper evaluates the approach only on two psychiatric classification tasks using region-of-interest features, which may limit generalizability to other neuroimaging modalities or clinical conditions. The method relies on prior neuroanatomical knowledge to define groups for whitening, which requires domain expertise and may not be straightforward for all brain structures or research questions. Additionally, the paper does not provide detailed comparisons with other interpretability-focused approaches beyond mentioning PCA and ICA, and the specific clinical validation of whether the improved interpretability actually translates to actionable neurobiological insights is not thoroughly demonstrated.
✨ Generated by Claude · Apr 25, 2026 · Read the PDF for authoritative content.