Investigating the effect of changing parameters when building prediction models for post-stroke aphasia

Area of Science:

• Neuroimaging
• Neurology
• Computational Neuroscience

Background:

• Neuroimaging advances understanding of brain-language mapping in healthy and impaired individuals, including post-stroke aphasia.
• Emerging research focuses on reverse inference: building brain-to-behavior prediction models from neuroimaging data.

Purpose of the Study:

• To investigate the impact of brain partitions, multimodal neuroimaging, and machine learning algorithms on predicting language and cognition in post-stroke aphasia.
• To establish principles for future neuroimaging-based outcome prediction in neurological patients.

Main Methods:

• Explored prediction models for four key dimensions of language and cognition in post-stroke aphasia.
• Evaluated the predictive performance using different brain partitions and machine learning algorithms.
• Compared models using structural MRI (T1 scans) with those using diffusion-weighted imaging data.

Main Results:

• Prediction models using diffusion-weighted data did not outperform models based on structural T1 scan measures across all four behavioral dimensions.
• The choice of brain partitions and machine learning algorithms influenced model performance.
• Structural MRI data provided robust features for predicting language and cognitive variations in aphasia.

Conclusions:

• Structural MRI measures are effective for predicting language and cognitive outcomes in post-stroke aphasia.
• Diffusion-weighted imaging does not offer additional predictive value over structural MRI for these specific outcomes.
• The study provides guidance on selecting neuroimaging features and methods for predicting neurological patient outcomes.