A Lesion-aware Edge-based Graph Neural Network for Predicting Language Ability in Patients with Post-stroke Aphasia.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Neurology

Background:

• Post-stroke aphasia significantly impacts communication and quality of life.
• Accurate prediction of language ability is crucial for effective rehabilitation strategies.
• Current methods may not fully capture the complex relationship between brain connectivity and language deficits.

Purpose of the Study:

• To introduce a lesion-aware graph neural network (LEGNet) for predicting language ability from resting-state fMRI (rs-fMRI) connectivity.
• To evaluate LEGNet's performance against baseline deep learning models in post-stroke aphasia patients.
• To assess LEGNet's generalization capabilities on independent datasets.

Main Methods:

• Developed LEGNet, integrating edge-based, lesion encoding, and subgraph learning modules.
• Utilized synthetic Human Connectome Project (HCP) data for model pretraining and hyperparameter tuning.
• Performed repeated 10-fold cross-validation on an in-house dataset of post-stroke aphasia patients.

Main Results:

• LEGNet significantly outperformed baseline deep learning methods in predicting language ability.
• The model demonstrated superior generalization performance on a second, independent in-house dataset.
• LEGNet effectively learned the intricate relationships between rs-fMRI connectivity and language function in lesioned brains.

Conclusions:

• LEGNet shows significant potential for enhancing the evaluation of language ability in post-stroke aphasia.
• The model's ability to integrate lesion information and functional connectivity offers a novel approach.
• This work highlights the utility of advanced machine learning for understanding brain-behavior relationships in neurological disorders.