Predicting Post-Stroke Aphasia Speech Performance from Multimodal Data with Explainable Machine Learning

Area of Science:

• Neuroscience
• Computational Linguistics
• Speech-Language Pathology

Background:

• Aphasia is a common post-stroke language impairment, often becoming chronic.
• Current prediction methods for aphasia recovery have limited accuracy.
• Personalized predictions are needed to optimize aphasia therapies.

Purpose of the Study:

• To predict word-by-word speech accuracy in persons with aphasia (PWA).
• To enable personalized speech therapies by improving prediction accuracy.
• To develop clinically applicable models using accessible inputs and explainable features.

Main Methods:

• Combined multimodal inputs: clinical scores, structural MRI neuroimaging, and word-by-word linguistic difficulty metrics (cognitive and articulatory burden).
• Utilized naturalistic corpora (>1 billion words) to compute linguistic difficulty.
• Employed retrospective training, cross-validation, and bootstrapping with random forest classifiers on 4620 trials.

Main Results:

• Multimodal models significantly outperformed single-input models (AUROC up to 0.90 ± 0.04).
• Key predictors included Western Aphasia Battery scores, semantic demands, word length (phonemes, syllables), and brain structural integrity.
• A simplified, clinically deployable model (AphasiaLENS) showed strong prospective generalization (AUROC 0.81-0.89).

Conclusions:

• Machine learning models integrating linguistic difficulty, clinical data, and neuroimaging can accurately predict PWA speech accuracy.
• A simplified, explainable model (AphasiaLENS) offers a clinically viable tool for personalized aphasia treatment planning.
• The findings enhance understanding of brain-behavior relationships in aphasia and guide future research targets.