How AI Supports Our Mission

This study compared Support Vector Regression (SVR) and Random Forest (RF) models trained on multimodal MRI data (i.e., structural, diffusion, and resting-state fMRI) from 76 individuals with post-stroke aphasia to predict aphasia severity. Using recursive feature elimination and nested cross-validation, the SVR model combining resting-state functional connectivity and white-matter integrity achieved the best performance, highlighting how integrating functional and structural data enhances precision in modeling language impairment severity.
 

Hu, X., Varkanitsa, M., Kropp, E., Betke, M., Ishwar, P., & Kiran, S. (2025). Aphasia severity prediction using a multi-modal machine learning approach. NeuroImage, 317, 121300. https://doi.org/10.1016/J.NEUROIMAGE.2025.121300

Read the Paper

• 

  Xinyi (Selena) Hu

  xhu07@bu.edu

This study introduced LEGNet, a lesion-aware graph neural network (GNN) that predicts language ability from resting-state fMRI in post-stroke aphasia. The model integrates three modules—edge-based learning, lesion encoding, and subgraph learning—to represent how stroke lesions reshape functional connectivity. Using synthetic lesion data from the Human Connectome Project for pretraining and repeated cross-validation on two in-house datasets, LEGNet outperformed traditional ML and deep learning baselines, including BrainGNN, BrainNetCNN, and SVR. The approach demonstrates how lesion-informed GNNs can model network-level reorganization and improve prediction of language outcomes after stroke.
 

Chen, Z., Varkanitsa, M., Ishwar, P., Konrad, J., Betke, M., Kiran, S., & Venkataraman, A. (2025). A Lesion-Aware Edge-Based Graph Neural Network for Predicting Language Ability in Patients with Post-stroke Aphasia. Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) , 15266 LNCS, 91–101. https://doi.org/10.1007/978-3-031-78761-4_9/FIGURES/3

Read the Paper

• 

  Zijian Chen

  zijianc@bu.edu

This study tested support vector machine (SVM) and random forest (RF) models to predict who benefits from 12 weeks of language therapy after stroke. Using pretreatment behavioral, demographic, and multimodal MRI data from 55 individuals, the SVM combining aphasia severity, demographics, white-matter integrity, gray-matter sparing, and resting-state connectivity achieved the highest accuracy. The results show that resting-state connectivity and structural integrity jointly drive treatment responsiveness, supporting AI-based precision rehabilitation.
 

Billot, A., Lai, S., Varkanitsa, M., Braun, E. J., Rapp, B., Parrish, T. B., Higgins, J., Kurani, A. S., Caplan, D., Thompson, C. K., Ishwar, P., Betke, M., & Kiran, S. (2022). Multimodal Neural and Behavioral Data Predict Response to Rehabilitation in Chronic Poststroke Aphasia. Stroke, 53(5), 1606–1614. https://doi.org/10.1161/STROKEAHA.121.036749

Read the Paper

• 

  Swathi Kiran

  kirans@bu.edu

This study used the BiLex neural network, built on Self-Organizing Maps (SOMs), to simulate how Spanish–English bilinguals with post-stroke aphasia respond to language therapy. Each patient’s model was “lesioned” to mimic brain injury and retrained to reproduce therapy outcomes in both languages. Using an evolutionary optimization algorithm and leave-one-out cross-validation, BiLex accurately predicted improvement in the treated language and captured degrees of cross-language transfer. These results demonstrate that computational patient models can forecast therapy response and inform language-specific treatment planning for bilingual aphasia.
 

Grasemann, U., Peñaloza, C., Dekhtyar, M., Miikkulainen, R., & Kiran, S. (2021). Predicting language treatment response in bilingual aphasia using neural network-based patient models. Scientific Reports, 11(1), 1–11. https://doi.org/10.1038/S41598-021-89443-6;SUBJMETA

Read the Paper

• 

  Swathi Kiran

  kirans@bu.edu

This NIH-funded project develops computational models that integrate multimodal neuroimaging with behavioral data to disentangle the complex mechanisms underlying aphasia recovery. Leveraging advances in artificial intelligence, we use biologically-informed neural networks tailored to each imaging modality and automated feature-selection strategies to isolate key drivers of recovery. The models are trained and validated on three longitudinal datasets, providing one of the most comprehensive multimodal evaluations of post-stroke aphasia to date and laying the groundwork for individualized, data-driven recovery prediction.
Learn More

• 

  Archana Venkataraman

  archanav@bu.edu

This ongoing research leverages natural language processing to automate the detection of language impairments such as aphasia and dementia. The pipeline extracts linguistic and semantic features using word embeddings from patient speech and uses trained machine learning models to distinguish clinical cases from healthy controls, aiding early diagnosis and intervention.

Cassie Lee seunghee@bu.edu