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

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Artificial Intelligence in Medicine

Background:

• Post-stroke aphasia significantly impacts communication and quality of life.
• Resting-state functional magnetic resonance imaging (rs-fMRI) reveals brain functional connectivity patterns.
• Accurate prediction of language ability is crucial for effective rehabilitation strategies.

Purpose of the Study:

• To develop and validate a novel deep learning model, LEGNet, for predicting language ability in post-stroke aphasia patients.
• To investigate the utility of integrating lesion information with rs-fMRI connectivity for enhanced prediction accuracy.
• To assess the model's performance and generalization capabilities on independent datasets.

Main Methods:

• Development of a lesion-aware graph neural network (LEGNet) incorporating edge-based learning, lesion encoding, and subgraph learning modules.
• Model pretraining and hyperparameter tuning using synthetic data from the Human Connectome Project (HCP).
• Evaluation via repeated 10-fold cross-validation on an in-house post-stroke aphasia dataset and testing on a second independent dataset.

Main Results:

• LEGNet significantly outperformed baseline deep learning methods in predicting language ability.
• The model demonstrated superior generalization performance on a dataset acquired with a different neuroimaging protocol.
• Integration of lesion information improved the model's ability to capture brain-language relationships.

Conclusions:

• LEGNet effectively learns the complex relationships between rs-fMRI connectivity, brain lesions, and language ability in aphasia patients.
• The proposed model offers a promising tool for objective and accurate post-stroke aphasia evaluation.
• This approach has the potential to guide personalized rehabilitation strategies and improve patient outcomes.