🤖 machine learning

Improving clinical interpretability of linear neuroimaging models through feature whitening

This paper introduces an anatomically informed feature whitening approach that decorrelates shared variance between brain regions to enhance the clinical interpretability of linear neuroimaging model weights while preserving predictive performance in psychiatric classification tasks.

Original authors: Sara Petiton, Antoine Grigis, Raphaël Vock, Edouard Duchesnay

Published 2026-04-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Sara Petiton, Antoine Grigis, Raphaël Vock, Edouard Duchesnay

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to figure out which ingredients in a complex soup are responsible for its specific taste. You have a recipe (a computer model) that tells you how much of each ingredient contributed to the final flavor. But here's the problem: in this soup, the ingredients are all mixed up and clinging to each other. For example, the amount of salt and the amount of pepper might always go up and down together because they are usually added at the same time.

If you ask the recipe, "How much did the salt contribute?" it might say, "A lot!" But it might also say, "And the pepper contributed a lot!" even though they are just two sides of the same coin. The recipe gets confused because it can't tell if the flavor comes from the salt, the pepper, or the fact that they are always paired together. This is exactly the problem scientists face when studying brain scans.

The Problem: The "Echo Chamber" of the Brain

In brain imaging, different parts of the brain are naturally connected.

The Left-Right Echo: The left and right sides of your brain (like the left and right amygdala) are like twins. They usually do things together. If the left side changes, the right side usually changes too.
The Volume Trade-off: Inside a specific brain region, if the "gray matter" (the brain's processing power) shrinks, the "cerebrospinal fluid" (the liquid surrounding it) often expands to fill the gap. They are inversely related.

When researchers use standard computer models to find biomarkers for diseases like Schizophrenia or Bipolar Disorder, these models get confused by these natural connections. The model assigns "weights" (importance scores) to brain regions, but because the regions are so correlated, the scores get muddled. It's like trying to hear a solo violinist in a room where every other instrument is playing the exact same note at the same volume. You can't tell who is actually leading the song.

The Solution: "Whitening" the Data

The authors of this paper introduced a clever trick called Feature Whitening.

Think of "whitening" like putting on noise-canceling headphones that don't just block noise, but actually rearrange the sound so every instrument has its own unique space.

The Separation: Instead of letting the left and right brain twins talk over each other, the method mathematically "uncouples" them. It asks the model: "If the left side did this, what would the right side have to do if they weren't linked?"
The Regularizer: They also added a "dimmer switch" (called regularization). Sometimes, you want to keep a little bit of the connection between the twins because that connection itself might be a sign of disease. The dimmer switch lets them control how much they separate the twins, keeping just enough of the relationship to be useful but not so much that it confuses the model.

The Experiment: Sorting the Soup

The researchers tested this on two groups of patients: those with Bipolar Disorder and those with Schizophrenia, comparing them to healthy people.

The Test: They ran the computer models twice. Once with the messy, mixed-up data (the original soup), and once with the "whitened" data (the rearranged soup).
The Result:
- Accuracy: The "whitened" model was just as good at predicting who was sick as the original model. It didn't lose any power.
- Clarity: This is where the magic happened. With the whitening, the model's "weights" suddenly made sense.
  - Before whitening, the model might have been confused about which brain part was important.
  - After whitening, the model pointed clearly to the specific brain regions known to be affected by these diseases (like the hippocampus or the ventricles), matching what other major studies have found.

The Takeaway

Imagine you are a detective trying to solve a crime. You have a list of suspects, but they all have alibis that look identical because they were all together at the scene. It's hard to know who actually committed the crime.

This paper introduces a new way to interrogate the suspects. By "whitening" the data, the researchers separated the suspects' stories. Suddenly, the computer model could clearly say, "It wasn't the group; it was this specific person (this specific brain region) who is the key to the case."

In short: They found a way to untangle the brain's natural "echoes" so that AI models can give doctors clearer, more trustworthy answers about which parts of the brain are actually involved in mental illness, without losing any accuracy in their predictions. It turns a confusing chorus into a clear solo performance.

1. Problem Statement

Linear models (e.g., logistic regression) are widely used in computational neuroimaging to identify biomarkers for psychiatric disorders. However, interpreting the learned model weights is often challenging due to collinearity (high correlation) between features.

The Core Issue: In neuroimaging, brain regions are inherently correlated. Specifically, homologous structures in the left and right hemispheres exhibit strong anatomical correlations, and Gray Matter (GM) and Cerebrospinal Fluid (CSF) volumes within the same region often have an inverse relationship.
Consequence: When features are correlated, linear model weights reflect shared variance rather than region-specific contributions. This obscures the biological meaning of the weights, making it difficult to link model outputs to specific neurobiological mechanisms or clinical insights.
Limitations of Existing Solutions: While "forward models" attempt to address collinearity post-hoc, they often ignore known brain structures (like inter-hemispheric correlations). Standard dimensionality reduction techniques like PCA or ICA use whitening but discard information, which is unsuitable for feature interpretation where the full input signal must be retained.

2. Methodology

The authors propose a feature-level whitening approach that decorrelates specific pairs of neuroanatomically informed features while retaining the full input dimensionality.

A. Data and Preprocessing

Datasets: Two psychiatric classification tasks were performed:
1. Bipolar Disorder (BD) vs. Healthy Controls (HC) using BIOBD and BSNIP datasets (861 participants).
2. Schizophrenia (SCZ) vs. HC using the SCHIZCONNECT-VIP dataset (604 participants).
Features: Voxel-based morphometry (VBM) measures extracted from T1-weighted MRI. The brain was parcellated into 140 Regions of Interest (ROIs), capturing both GM and CSF volumes, resulting in 280 total features.
Model: Logistic Regression was chosen for its linearity and interpretability.

B. The Whitening Approach (ZCA-cor)

The authors utilize Zero-Phase Component Analysis with Correlation (ZCA-cor) whitening, applied pairwise to specific feature groups:

Left-Right Pairs: Homologous regions (e.g., Left Amygdala GM vs. Right Amygdala GM).
GM-CSF Pairs: Tissue types within the same region (e.g., Left Amygdala GM vs. Left Amygdala CSF).

Mathematical Formulation:
For a pair of standardized features $X_p$ , the correlation matrix $R_p$ is decomposed via Singular Value Decomposition (SVD): $R_p = U_p \Lambda_p U_p^\top$ .
The whitening matrix $W_p$ is defined as:
$W_p = U_p \Lambda_p^{-1/2} U_p^\top$
The whitened features are $Z_p = X_p W_p$ , ensuring $Corr(Z_p) = I$ (identity matrix).

C. Regularized Whitening

To preserve potentially disease-relevant correlations (e.g., if inter-hemispheric coupling differs between patients and controls), the authors introduce a regularized variant:
$W_\alpha = \alpha W + (1 - \alpha)I$
Where $\alpha \in [0, 1]$ controls the degree of decorrelation.

$\alpha = 1$ : Full whitening.
$\alpha < 1$ : Partial whitening, retaining some original correlation structure.

D. Weight Projection

Crucially, the classifier is trained on the whitened data ( $Z$ ), but the weights are projected back to the original feature space ( $X$ ) to maintain interpretability:
$\theta = W_{ZCA-cor}^\top \beta$
Where $\beta$ are the weights in the whitened space and $\theta$ are the interpretable weights in the original space. The paper proves that this transformation preserves the relative ordering of feature contributions (i.e., if feature A is more important than B in the whitened space, it remains so in the original space).

3. Key Contributions

Novel Application: First application of ZCA-cor whitening specifically to improve the interpretability of linear neuroimaging models, distinct from its traditional use in dimensionality reduction or noise removal.
Anatomically Informed Decorrelation: Unlike generic whitening, this method targets specific, biologically meaningful pairs (hemispheres and tissue types) based on prior neuroanatomical knowledge.
Regularized Variant: Introduction of a tunable parameter ( $\alpha$ ) to balance decorrelation with the preservation of disease-specific covariance patterns.
Preservation of Signal: Unlike PCA/ICA, this method retains the full input signal, making it suitable for feature selection and interpretation rather than just compression.

4. Results

The method was evaluated on BD and SCZ classification tasks using 10-fold cross-validation.

Predictive Performance: Whitening did not degrade classification performance.
- BD vs. HC: ROC-AUC remained stable (Original: 76.39% vs. Whitened: 76.24%).
- SCZ vs. HC: ROC-AUC remained stable (Original: 81.68% vs. Whitened: 81.0%).
- Statistical tests showed no significant difference ( $p > 0.05$ ).
Feature Disentanglement:
- Correlation matrices showed a marked reduction in correlations between homologous left/right regions and between GM/CSF volumes after whitening.
- Feature dispersion became more spherical, indicating successful decorrelation.
Enhanced Interpretability:
- The model weights derived from whitened features showed higher alignment with established neurobiological literature (specifically ENIGMA meta-analyses).
- BD: Whitening strengthened correlations with cortical rankings and elevated subcortical regions (ventricles, hippocampus, thalamus, amygdala) to match known findings.
- SCZ: Improved the ranking of key regions (hippocampus, inferior temporal gyrus, fusiform gyrus, lateral ventricles) compared to the unwhitened model.
- The top contributing features in the whitened models were more biologically plausible than those in the original models.

5. Significance and Conclusion

This work provides a robust framework for linking linear model outputs to neurobiological mechanisms in psychiatry. By decoupling shared variance from region-specific contributions, the proposed method allows clinicians and researchers to trust model weights as more accurate indicators of specific brain pathology.

Clinical Impact: The approach bridges the gap between machine learning predictions and clinical understanding, offering a tool to identify more reliable biomarkers for Bipolar Disorder and Schizophrenia.
Future Directions: While currently applied to linear models and ROI features, the authors suggest this methodology could be extended to voxel-wise data and integrated into complex deep learning architectures to enhance interpretability in broader neuroimaging contexts.