Strategies for feature extraction from structural brain imaging in lesion-deficit modelling

Area of Science:

• Cognitive Neuroscience
• Neuroimaging
• Machine Learning

Background:

• High-dimensional modeling of post-stroke deficits using brain imaging is crucial for understanding cognitive neuroscience and guiding rehabilitation.
• Optimizing model performance relies on effective feature selection and representation strategies.
• The comparative performance of different feature representation methods in lesion-deficit modeling remains unclear.

Purpose of the Study:

• To compare the performance of three common feature representation strategies in multivariate lesion-deficit models.
• To evaluate voxel-wise, lesion-anatomical componential reduction, and region-wise atlas-based representations.
• To determine the optimal strategy for predicting post-stroke deficits from structural brain imaging data.

Main Methods:

• Utilized multivariate, machine-learning-based lesion-deficit models to predict stroke-related deficits.
• Employed support vector regression, optimized via nested cross-validation.
• Compared voxel-wise, lesion-anatomical, and atlas-based feature representations using structural lesion data.
• Validated model performance on held-out data.

Main Results:

• Numerically superior models were consistently achieved with lower-dimensional, featurized data, particularly principal components from lesion maps.
• Minor, non-significant performance differences were observed across the three feature representation strategies.
• Model performance did not significantly differ between various popular brain atlases.
• Fine-tuning feature representations may offer minor benefits, but lesion data complexity poses challenges.

Conclusions:

• All three common feature representation strategies (voxel-wise, lesion-anatomical, atlas-based) are generally suitable for lesion-deficit modeling.
• The choice of feature representation strategy has a limited impact on model performance.
• Understanding the complex interplay of lesion-anatomical and functional criteria is essential for effective feature reduction in stroke research.