Stable multivariate lesion symptom mapping.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Medical Image Analysis

Background:

• Multivariate lesion-symptom mapping (MLSM) analyzes whole-brain lesions to predict neurological impairments.
• High-dimensional feature spaces in MLSM can be challenging due to irrelevant or redundant features.
• Distinguishing critical brain features for behavior prediction requires robust methods.

Purpose of the Study:

• Introduce stable multivariate lesion-symptom mapping (sMLSM) integrating feature reliability and stability selection.
• Develop and provide an open-source MATLAB implementation of sMLSM.
• Enhance the accuracy and interpretability of lesion-symptom mapping for predicting neurological impairments.

Main Methods:

• Implemented sMLSM by incorporating stability selection into conventional MLSM.
• Utilized open-source MATLAB code for analysis.
• Validated sMLSM on two independent public stroke datasets (N=167 chronic, N=1106 acute).

Main Results:

• sMLSM effectively eliminates inconsistent features and reduces feature weight variation compared to MLSM.
• The method improves prediction accuracy for aphasia severity, outperforming lesion size alone, even with moderate sample sizes (N>75).
• sMLSM reveals a distinct spatial distribution of feature importance, highlighting both univariate and MLSM-identified regions.

Conclusions:

• sMLSM offers a more robust and accurate approach to lesion-symptom mapping.
• The method enhances the identification of reliable neural biomarkers for neurological impairments.
• sMLSM provides deeper insights into the neurobiology of brain-behavior relationships.