Simulating Bilingual Aphasia Rehabilitation: Evidence from a computational model

Swathi Kiran, Uli Grasemann, Chaleece Sandberg, Risto MiikkulainenThe Aphasiology Archive (University of Pittsburgh) 2011

TL;DR

Researchers have developed a computational model called DISLEX that uses self-organizing maps to simulate how bilingual aphasia affects language and how patients recover. By "lesioning" the model with digital noise, they successfully replicated complex naming impairment patterns and demonstrated how rehabilitation in one language can trigger recovery in another—a phenomenon known as cross-language transfer.

What Problem Does This Paper Address?

When a bilingual person suffers a stroke resulting in aphasia (language impairment), the road to recovery is often a "black box." Clinicians struggle to determine the optimal rehabilitation strategy because every patient’s pre-stroke background is different. Factors like Age of Acquisition (AoA)—when they learned each language—and their relative proficiency are usually assessed subjectively after the brain damage has already occurred.

The researchers wanted to move beyond subjective estimates. They sought to create a mechanistic, objective tool to understand how language is represented in the brain, how stroke-induced damage disrupts those representations, and why training in one language (e.g., English) sometimes helps a patient improve in their other language (e.g., Spanish).

Methods

The authors utilized a connectionist approach called DISLEX, based on the "Revised Hierarchical Model" of bilingualism.

  • Architecture: The model consists of three Self-Organizing Maps (SOMs): one for semantics (meanings) and two for phonetics (the sounds/words of L1 and L2). Each map contains a 30x40 grid of neurons.
  • Training: The maps were trained simultaneously using Gaussian neighborhood functions. To mimic human development, the "neighborhood" size decreased over time, allowing the model to form a global structure early on and refine local similarities later.
  • Simulating Variables:
    • AoA: For "late" learners, training on the L2 map started after the L1 map was already established.
  • Proficiency: This was simulated by varying the amount of exposure (the number of word pairs) the model received in each language.
  • The "Lesion": To simulate a stroke, the researchers added Gaussian noise to the connections between the semantic map and the phonetic maps. They varied the "lesion strength" (the standard deviation of the noise) to see how performance degraded.
  • Rehabilitation: The "damaged" model was retrained in one language until it reached 90% accuracy, while researchers monitored the untrained language for improvements.

Datasets

The study didn't use a massive external database but rather focused on matching the model to specific human profiles:

  • Normal Bilinguals (N=39): Data from healthy speakers with varying AoAs and proficiencies were used to calibrate the model's training parameters.
  • Aphasia Patients (N=19): The model attempted to replicate the self-reported pre-stroke performance and post-stroke naming deficits of 19 actual patients.

Key Insights and Ideas

  • Symmetric Damage, Asymmetric Results: One of the most striking findings was that even if the "stroke" (noise) was applied equally to both languages, the resulting impairment was often asymmetric. This happens because the language with lower pre-stroke proficiency is less "robust" and collapses faster under the same amount of neural noise.
  • The Power of Transfer: The model successfully simulated cross-language transfer (CLT). In a scenario where a patient had high Spanish exposure but low English exposure, training them in English (the weaker language) actually boosted their Spanish performance.
  • Modeling Subjectivity: The model revealed inconsistencies in human data. In cases where the model couldn't match a patient's performance, the authors suggest it might be because the patients' self-reports of their pre-stroke abilities were inaccurate—highlighting the model's potential as a more objective "sanity check" for clinical history.

Key Results

  • Patient Matching: The DISLEX model successfully matched the naming patterns of 13 out of 19 aphasia patients.
  • Subgroup Identification: The researchers identified three distinct patterns of impairment:
  • Symmetric damage leading to asymmetric impairment (due to proficiency differences).
  • Asymmetric damage leading to asymmetric impairment.
  • Symmetric damage leading to symmetric impairment.
  • Rehab Success: Preliminary results showed that retraining the model in one language (reaching 90% accuracy) consistently led to improvements in the untrained language, mirroring real-world clinical observations from previous studies (e.g., Edmonds & Kiran, 2006).

Limitations and Future Work

The authors note that the rehabilitation component of the project is still ongoing. While the preliminary results are promising, they have yet to explore the full range of variables that might influence recovery.

Additionally, the model relies on the accuracy of the "pre-stroke" parameters. If the initial training of the model doesn't perfectly capture a patient's unique linguistic history, the simulation of their recovery may be flawed. Future work will likely involve refining these parameters and exploring more complex "lesion" sites within the network to better mimic the diverse nature of brain injuries.

Summary updated

Abstract

Current research on bilingual aphasia has only begun to inform us about optimal rehabilitation but the literature is sparse in terms of interpreting the nature of naming impairments in bilingual aphasia. The goal of the present study is to build a model that can simulate language impairments and rehabilitation in patients with bilingual aphasia. Specifically, the model takes into account Age of Acquisition (AoA) and pre-stroke language use in the two languages. This model is subsequently lesioned at specific sites and then rehabilitated in one language (e.g., either English or Spanish) and the extent of cross language transfer is examined.
BibTeX
@inproceedings{kiran2011simulating,
  author = {Swathi Kiran and Uli Grasemann and Chaleece Sandberg and Risto Miikkulainen},
  title = {{Simulating Bilingual Aphasia Rehabilitation: Evidence from a computational model}},
  booktitle = {The Aphasiology Archive (University of Pittsburgh)},
  year = {2011},
}

Loading related content...

Figure