Predicting Aphasia Interpersonal Communication Performance from Clinically Available Multimodal Data using Machine Learning
Shreya Parchure1, Arnav Gupta1, Leslie Vnenchak1, Olufunsho Faseyitan1, Apoorva Kelkar2, Denise Harvey1, John Medaglia2, Harry Coslett1, Roy Hamilton1
1Neurology, University of Pennsylvania, 2Drexel University
Objective:
Utilize routinely available clinical data from Persons with Aphasia (PWA) to predict detailed interpersonal communication with >80% accuracy (current gold standard). Employ explainable machine learning (ML) to identify key features influencing personalized language impairment. Generalize models to unseen PWA across etiologies of language impairment from Stroke to Alzheimer's Disease.
Background:
Clinical predictions of language performance in Persons with Aphasia (PWA) are limited to insights from lesion size, location, and patient demographics; having limited accuracy and specificity. Unfortunately, novel research algorithms have had limited clinical value because they rely on clinically unavailable data (e.g., fMRI), and focus on classifying patient severity rather than predicting language task impairments that can inform treatments.
Design/Methods:
We trained and validated random forest ML classifiers for noun accuracy using 6200 interpersonal communicative exchanges from 52 PWA (WAB-AQ 25-80, 41 chronic post-Stroke Aphasia, 11 Alzheimer’s Disease). Inputs included: 1) Clinical characteristics (language and cognitive testing, lesion volume, demographics); 2) White matter connectivity from DSI MRI across 83 cortical regions; and 3) Linguistic difficulty (semantic selection, phonemes, etc.) computed for any given word using naturalistic speech corpora. Classifiers were optimized using cross-validation and evaluated on new PWA and unseen language prompts.
Results:
The best classifiers (AUC=0.9-0.89) combined linguistic difficulty with brain networks or clinical testing. They significantly (p<0.05) outperformed single-datasets or combinations without linguistics (AUC=0.83-0.69). Key predictors included word length and retrieval difficulty, WAB-AQ sub-scores, and white matter connectivity of intact and differentially lesioned brain regions (e.g., left insula and inferior temporal gyri) obtained from structural MRI scans. Prospective generalization to new PWA, unseen language prompts, and translation from Stroke to Alzheimer's retained statistically similar (p>0.05) performance of the best model using all input data (AUC=0.85-0.93).
Conclusions:
Our explainable ML model effectively predicts interpersonal communication performance in PWA using multimodal clinically available data. This approach generalizes to new patients, prompts, and etiologies.
Disclaimer: Abstracts were not reviewed by Neurology® and do not reflect the views of Neurology® editors or staff.