Machine learning predictions of recovery in bilingual post-stroke aphasia: Aligning insights with clinical evidence
Authors: Manuel Marte, Erin Carpenter, Michael Scimeca, Marissa Russell-Meill, Claudia Peñaloza, Uli Grasemann, Risto Miikkulainen, and Swathi Kiran, Accepted in the Journal Stroke
Q&A with Manuel Marte
What is this paper about?
The paper examines how well machine learning models can predict language recovery outcomes in bilingual stroke survivors with aphasia. The study analyzed data from 48 Spanish-English bilingual individuals who received language therapy, looking at factors like aphasia severity, education level, and cognitive abilities to predict who would improve after treatment and whether improvements in one language would transfer to the other language.
The results are important for two key reasons. First, the machine learning models were relatively successful at predicting treatment outcomes, which could help clinicians make more informed decisions about therapy approaches for bilingual patients. Second, the study identified specific factors that matter most for recovery – particularly that the severity of language impairment and cognitive abilities at baseline were strong predictors of improvement. This is especially significant given the growing Hispanic population in the United States and the higher risk of stroke in this community. Advancing an understanding of these predictive factors will hopefully lead to more personalized and effective treatment plans for bilingual stroke survivors.
How do the findings relate to the brain and recovery?
The study relates to brain and recovery in a few different ways. First, the study revealed that baseline language abilities strongly predict significant recovery. Aphasia severity relates closely to the neurobiological integrity of language in the brain. When a person had less severe aphasia in the treated language, they were more likely to improve with therapy. Therefore, this suggests that having more preserved neural networks provides better “scaffolding” for treatment-induced recovery of language skills.
Second, the study found that cognitive abilities play a crucial role in recovery. Better performance on cognitive tests predicted better treatment outcomes, indicating that general cognitive resources support language recovery.
Third, the findings about cross-language transfer (i.e., improvements in the untreated language) tell us about how bilingual language systems interact during recovery. The fact that improvement in one language can and do generalize to the other (in some individuals) suggests that shared representations between languages remain accessible after stroke.
In sum, the clinical implications are that assessment of both language and cognitive abilities is crucial for predicting recovery potential and that therapy may be most effective when it can leverage preserved abilities in either language to support overall recovery.
Background: Predicting treated language improvement (TLI) and transfer to the untreated language (cross-language generalization, CLG) after speech-language therapy in bilingual individuals with post-stroke aphasia is crucial for personalized treatment planning. This study evaluated machine learning (ML) models to predict TLI and CLG, and identified the key predictive features (e.g., patient severity, demographics, and treatment variables) aligning with clinical evidence.
Methods: 48 Spanish-English bilingual individuals with post-stroke aphasia received 20 sessions of semantic feature-based naming treatment in either their first or second language. Comprehensive language, cognitive, and background bilingual experience assessments were administered pre- and post-treatment. Sixteen curated features spanning demographics, language abilities, cognition, and bilingual experience were used as inputs to six ML algorithms to predict treatment responders vs. non-responders and CLG vs no CLG.
Results: The top two ML models achieved F1 scores of 0.767 ± 0.153 for TLI and 0.790 ± 0.172 for CLG. Interpretability analyses revealed that aphasia severity in the trained language, education, and cognitive performance were key predictors of TLI. Aphasia severity in the untreated language and cognitive performance emerged as influential features of CLG. These aligned with expectations based on prior literature.
Conclusions: For the first time, ML models reveal that factors such as patient severity and demographics predict TLI and CLG after therapy in Spanish-English bilingual individuals with post-stroke aphasia. Consideration of both treated and untreated language severity, as well as cognitive assessment performance, when forecasting treatment outcomes in an underserved population such Spanish-English stroke survivors, can meaningfully impact their short-term and long-term clinical care.