Predicting language outcome after stroke using machine learning: in search of the big data benefit

Area of Science:

• Neuroscience
• Computational Linguistics
• Medical Informatics

Background:

• Accurate prediction of post-stroke language outcomes is crucial for effective rehabilitation in aphasia patients.
• Machine learning (ML) models offer potential for predicting these outcomes using neuroimaging and clinical data.

Purpose of the Study:

• To investigate the impact of sample size on the performance of logistic regression and deep learning (ResNet-18) models for predicting post-stroke language outcomes.
• To evaluate model performance on Spoken Picture Description and Naming tasks using the Comprehensive Aphasia Test.

Main Methods:

• Utilized data from 758 English-speaking stroke patients from the PLORAS project.
• Employed a learning curve approach to assess model performance with varying sample sizes.
• Applied Principal Component Analysis (PCA) to reduce neuroimaging data dimensionality.

Main Results:

• Logistic regression and deep learning models showed comparable performance, with accuracy plateauing around 80% for datasets larger than 300 patients.
• PCA indicated that neuroimaging data dimensionality could be significantly reduced (to 20 or even 2 components) without substantial accuracy loss.
• Results suggest classification may be driven by simpler patterns like lesion size, and current dataset sizes may limit further accuracy gains.

Conclusions:

• Current machine learning models show potential for predicting aphasia outcomes, but performance gains may be limited by existing dataset sizes.
• Larger datasets are needed to capture more complex patterns and potentially reach higher classification accuracy ceilings.
• Findings inform the practical application of ML in aphasia research and highlight the need for data aggregation to advance predictive modeling.