Benchmarking machine learning models in lesion-symptom mapping for predicting language outcomes in stroke survivors

Area of Science:

• Neuroscience
• Neurology
• Computational Neuroscience

Background:

• Decades of research link stroke-induced brain damage to language impairments.
• Traditional lesion-symptom mapping (LSM) studies often use mass univariate statistics, overlooking complex variable relationships.

Purpose of the Study:

• To benchmark machine learning (ML) models for predicting aphasia severity and naming impairment in chronic stroke survivors.
• To evaluate the efficacy of different neuroimaging modalities, parcellation schemes, and ML techniques.

Main Methods:

• Utilized structural and functional MRI data from 238 chronic stroke survivors for aphasia severity and 191 for naming impairment.
• Employed nested cross-validation to assess ML model performance across various parcellation atlases (JHU, AAL, BRO, AICHA), neuroimaging modalities (connectivity, diffusivity, anisotropy, lesion location), and ML algorithms (Random Forest, SVR, etc.).

Main Results:

• The optimal combination involved the JHU atlas, lesion location, and the Random Forest model.
• This combination demonstrated moderate to high correlations with behavioral scores for aphasia severity and naming impairment.
• Identified key brain regions within the language network, including perisylvian areas and associated pathways.

Conclusions:

• Machine learning offers a powerful complement to traditional LSM methods for understanding stroke-related language deficits.
• The study highlights the potential of integrating ML with neuroimaging and lesion data to enhance prediction of language recovery after stroke.