Enhanced estimations of post-stroke aphasia severity using stacked multimodal predictions

Area of Science:

• Neuroscience
• Neurology
• Medical Imaging

Background:

• Post-stroke aphasia severity and recovery are highly variable and difficult to predict.
• Optimal prediction requires integrating multiple neuroimaging modalities and multivariate brain-behavior analyses.

Purpose of the Study:

• To develop and validate a multimodal neuroimaging framework (STAMP) for predicting post-stroke aphasia severity.
• To assess the framework's accuracy compared to unimodal predictions and traditional methods.

Main Methods:

• Created a stacked multimodal prediction (STAMP) model using lesion maps, structural connectivity, and functional connectivity data.
• Crossvalidated predictions for four aphasia scores in 53 chronic stroke patients.
• Compared STAMP predictions with unimodal predictions and voxel-based lesion-to-symptom maps.

Main Results:

• STAMP achieved accurate crossvalidated predictions for all four aphasia scores (r=0.79-0.88).
• Significant prediction accuracy (r=0.66) was maintained in a full split crossvalidation.
• Multimodal predictions significantly outperformed any single modality alone.
• Topological maps showed high spatial congruency with traditional lesion-to-symptom mapping.

Conclusions:

• Neuroimaging modalities provide complementary information for predicting aphasia.
• A shift towards multimodal neuroimaging and multivariate methods is crucial for clinical translation.
• The STAMP framework offers a promising approach for predictive brain mapping in aphasia.