Predicting language treatment response in bilingual aphasia using neural network-based patient models

Uli Grasemann, Claudia Peñaloza, Maria Dekhtyar, Risto Miikkulainen, Swathi KiranScientific 2021 · DOI

Abstract

Predicting language therapy outcomes in bilinguals with aphasia (BWA) is complex due to varied pre- and post-stroke factors. This study utilizes BiLex, a computational model, to simulate language deficits and treatment responses in 13 Spanish-English BWA. By creating individualized "digital twins," the model accurately predicted treatment effects in the treated language and captured cross-language generalization. These findings suggest that computational modeling can guide personalized rehabilitation plans for BWA.

The Challenge of Bilingual Aphasia

Aphasia, a language disorder often resulting from stroke, presents unique challenges in bilingual individuals. Patients frequently exhibit asymmetrical impairment and recovery across their two languages. A primary clinical hurdle is identifying which language to target during therapy to maximize recovery in both—a phenomenon known as cross-language generalization. Because every patient possesses a unique "linguistic fingerprint" shaped by their history of language acquisition and exposure, standardized treatment protocols often fall short.

The BiLex Computational Framework

To address this variability, researchers developed BiLex, a neural network-based "patient model" utilizing Self-Organizing Maps (SOMs). SOMs mimic the brain's topological organization, where related concepts or sounds are grouped together.

The BiLex model architecture (a) and the organization of the semantic map (b, c).

Model Architecture and Simulation

The BiLex architecture consists of three interconnected maps: a shared semantic map representing word meanings and two language-specific phonetic maps for English and Spanish. The modeling process involves three key stages:

  1. Pre-stroke Training: The model is trained using a patient’s reported lifetime language exposure to establish a healthy bilingual baseline.
  2. Post-stroke Lesioning: Researchers "lesion" the model by removing neurons and connections until its performance matches the patient’s actual diagnostic scores on naming and semantic tests.
  3. Therapy Simulation: The damaged model is retrained on a subset of words, simulating a semantic-based therapy protocol. An Evolutionary Algorithm (EA) optimizes learning parameters to best fit individual patient data.

Key Results: Matching Deficits and Predicting Recovery

The BiLex model demonstrated exceptional accuracy in replicating post-stroke states and predicting therapy outcomes:

  • Impairment Matching: The model achieved $R^2$ values of 0.956 for semantic tests and 0.974–0.989 for naming tests across both languages.
  • Treatment Prediction: There was a highly significant correlation between predicted and actual recovery in the treated language ($p \ll 0.0001$).
  • Cross-Language Generalization: The model successfully captured "spillover" effects to the untreated language. Because the languages share a semantic map, retraining in one language can benefit the other through associative connections.
  • Robustness and Generalization: Leave-one-out cross-validation confirmed that BiLex could predict outcomes for new patients, a critical requirement for clinical utility.

Clinical Implications and Future Directions

The success of BiLex in simulating "digital twins" offers a promising path toward personalized medicine in speech-language pathology. By accounting for multi-level damage (semantic and phonetic), the model provides a more holistic view of aphasia than simpler models.

Future work aims to expand the model's vocabulary beyond concrete nouns and validate its predictive power in larger, prospective clinical trials. The ultimate goal is to use BiLex proactively to "prescribe" the optimal language for treatment, ensuring the most effective recovery path for every bilingual patient.

Summary updated

Abstract

Predicting language therapy outcomes in bilinguals with aphasia (BWA) remains challenging due to the multiple pre- and poststroke factors that determine the deficits and recovery of their two languages. Computational models that simulate language impairment and treatment outcomes in BWA can help predict therapy response and identify the optimal language for treatment. Here we used the BiLex computational model to simulate the behavioral profile of language deficits and treatment response of a retrospective sample of 13 Spanish-English BWA who received therapy in one of their languages. Specifically, we simulated their prestroke naming ability and poststroke naming impairment in each language, and their treatment response in the treated and the untreated language. BiLex predicted treatment effects accurately and robustly in the treated language and captured different degrees of cross-language generalization in the untreated language in BWA. Our cross-validation approach further demonstrated that BiLex generalizes to predict treatment response for patients whose data were not used in model training. These findings support the potential of BiLex to predict therapy outcomes for BWA and suggest that computational modeling may be helpful to guide individually tailored rehabilitation plans for this population.
BibTeX
@inproceedings{grasemann2021predicting,
  author = {Uli Grasemann and Claudia Peñaloza and Maria Dekhtyar and Risto Miikkulainen and Swathi Kiran},
  title = {{Predicting language treatment response in bilingual aphasia using neural network-based patient models}},
  booktitle = {Scientific},
  year = {2021},
  doi = {10.1038/s41598-021-89443-6},
}
Figure