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

High‐dimensional modelling of post‐stroke deficits from structural brain imaging is highly relevant to basic cognitive neuroscience and bears the potential to be translationally used to guide individual rehabilitation measures. One strategy to optimise model performance is well‐informed feature selection and representation. However, different feature representation strategies were so far used, and it is not known what strategy is best for modelling purposes. The present study compared the three common main strategies: voxel‐wise representation, lesion‐anatomical componential feature reduction and region‐wise atlas‐based feature representation. We used multivariate, machine‐learning‐based lesion‐deficit models to predict post‐stroke deficits based on structural lesion data. Support vector regression was tuned by nested cross‐validation techniques and tested on held‐out validation data to estimate model performance. While we consistently found the numerically best models for lower‐dimensional, featurised data and almost always for principal components extracted from lesion maps, our results indicate only minor, non‐significant differences between different feature representation styles. Hence, our findings demonstrate the general suitability of all three commonly applied feature representations in lesion‐deficit modelling. Likewise, model performance between qualitatively different popular brain atlases was not significantly different. Our findings also highlight potential minor benefits in individual fine‐tuning of feature representations and the challenge posed by the high, multifaceted complexity of lesion data, where lesion‐anatomical and functional criteria might suggest opposing solutions to feature reduction.