Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Machine Learning-based Prediction of Real-world Communication Accuracy in Post-Stroke Aphasia using Structural MRI, Clinical Aphasia Testing, and Linguistic Data as Key Predictors
Poster Session B, Friday, October 25, 10:00 - 11:30 am, Great Hall 3 and 4
Shreya Parchure1, Leslie Vnenchak1, Apoorva Kelkar2, Olufunsho Faseyitan1, John Medaglia2, Denise Harvey1, Roy Hamilton1, H. Branch Coslett1; 1University of Pennsylvania, 2Drexel University
Predictors of language performance in Persons with Aphasia (PWA) beyond general measures such as lesion size and location are currently lacking. Recent studies have sought to combine multiple neuroimaging modalities and behavioral testing to develop predictions. A limiting factor in these models is their reliance on data (e.g., fMRI) not typically available in clinical settings. To address this issue, we used machine learning approaches to predict performance of PWA during interpersonal communication in Constraint Induced Language Therapy (CILT). The algorithms were created to be explainable and to use predictors likely to be available in routine clinical settings. We also test the importance of various structural MRI neuroimaging features in the AI models, to depict how stroke damage and remaining connectivity of intact brain tissue translate into language function in aphasia. Participants included 40 individuals with chronic poststroke aphasia with WAB-AQ scores between 25-80. Data came from CILT session interactions involving a participant and therapist. Response accuracies of agent nouns were scored by blinded raters. Random forest classifiers were constructed to predict noun accuracy in CILT exchanges, using PWA stroke demographics, structural MRI white matter strengths, and linguistic priors about task difficulty. Data included 5472 trials from 40 PWA (33% of trials were correct and 66% incorrect), of which 90% was used for training classifiers. Models were optimized for the following parameters: number of independent decision trees making parallel classifications, and maximum depth of decision splits per tree. Multi-collinear features were hierarchically clustered using Ward’s linkage, to avoid erroneous importance of correlated neuroimaging regions. Feature importance was measured for each predictor in the model using 20 permutations and compared to a random number feature as null model benchmark. The best prediction of language performance was achieved by random forest trained on PWA stroke type, aphasia severity, linguistic metrics of task difficulty, and white matter connectivity strength from structural MRI (F1=0.89). This model resulted in a significantly superior performance compared with models trained on all features (F1=0.78, p<0.01) or single feature sub-sets of only stroke demographics, linguistic priors, or structural neuroimaging (F1 range=0.64–0.80, p<0.01). Of note, the best model showed high specificity for incorrect responses (93% recall score) with lower sensitivity for correct responses (77% precision). Important predictors in the model include linguistic priors (noun frequency and naming agreement of the trial), stroke severity (WAB-AQ score and aphasia type- especially whether Broca’s aphasia) and white matter strength of brain regions from structural MRI. Key predictor regions were: left insula, bilateral inferior temporal gyri, superior frontal regions, and deep gray matter structures such as thalamus and hippocampus. Using commonly available information from PWA after stroke, our model can accurately predict language performance for naming during interpersonal communication in a therapy setting. Important predictors are linguistic priors, structural neuroimaging connectivity, and aphasia characteristics; these factors carry complementary information in predicting language performance in PWA.
Topic Areas: Disorders: Acquired, Computational Approaches