Aphasia severity prediction using a multi-modal machine learning approach.

Area of Science:

• Neuroscience
• Medical Imaging
• Computational Linguistics

Background:

• Aphasia, a language disorder, significantly impacts stroke survivors.
• Predicting aphasia severity is crucial for effective rehabilitation planning.
• Current prediction models often rely solely on lesion characteristics.

Purpose of the Study:

• To investigate the efficacy of integrated neuroimaging modalities for predicting aphasia severity.
• To compare the predictive performance of Support Vector Regression (SVR) and Random Forest (RF) models.
• To identify key neuroimaging features that predict language outcomes in post-stroke aphasia.

Main Methods:

• Utilized T1 structural MRI, Diffusion Tensor Imaging (DTI), and resting-state fMRI (rsFMRI) data from 76 post-stroke aphasia patients.
• Employed SVR and RF models with supervised feature selection and stacked prediction.
• Assessed prediction accuracy using Western Aphasia Battery-Revised Aphasia Quotient (WAB-R AQ).

Main Results:

• The SVR model demonstrated superior performance (RMSE: 16.38±5.57, r: 0.70±0.13) over the RF model (RMSE: 18.41±4.34, r: 0.66±0.15).
• Resting-state neural activity and structural integrity were identified as critical predictors.
• Functional connectivity in bilateral homologous language areas significantly contributed to prediction accuracy.

Conclusions:

• Integrating multiple neuroimaging modalities enhances aphasia severity prediction beyond lesion information.
• Resting-state functional connectivity and structural integrity are vital for predicting language outcomes.
• Multi-modal neuroimaging approaches can inform personalized rehabilitation strategies for aphasia.