Presentation
Search Abstracts | Symposia | Slide Sessions | Poster Sessions | Lightning Talks
Discourse- and lesion-based aphasia severity estimation
There is a Poster PDF for this presentation, but you must be a current member or registered to attend SNL 2023 to view it. Please go to your Account Home page to register.
Poster A10 in Poster Session A, Tuesday, October 24, 10:15 am - 12:00 pm CEST, Espace Vieux-Port
Nicholas Riccardi1, Satvik Nelakuditi1, Chris Rorden1, Julius Fridriksson1, Rutvik H. Desai1; 1University of South Carolina
Introduction: Measuring aphasia severity, such as with the Western Aphasia Battery (WAB), can be burdensome on patients, their families, and clinicians. There is a need for supplementary measurements that are brief and can be administered remotely or by non-specialists. Such a measurement could be used to triage patients, monitor progress over time, or when patients have limited access to transportation. Discourse analysis, wherein patients produce continuous speech in response to a standardized prompt, provides a promising avenue towards these goals. However, the specific relationships between linguistic elements elicited by discourse and aphasia severity remain understudied. Here, we used machine learning to predict aphasia severity (WAB Aphasia Quotient; AQ) from linguistic features elicited by discourse prompts developed by AphasiaBank. We inspected feature weights from the models to draw conclusions about which linguistic features are most important for predicting AQ. We also supplemented and compared discourse-based models with lesion-based models that use structural neuroimaging features. Methods: 71 survivors of stroke (left hemisphere unilateral, > 6 months post-stroke) were recorded while responding to 3 AphasiaBank discourse prompts: picture-sequence (Broken Window; BW), narrative (Cinderella; ‘tell me the story of Cinderella’), and procedural (PBJ; ‘tell me how to make a peanut-butter and jelly sandwich). Recordings were transcribed and coded by trained research assistants under the supervision of speech-language pathologists (SLP). Computerized Language Analysis (CLAN) software extracted 45 linguistic discourse features (e.g., # of nouns, mean utterance length, etc.) for each patient and prompt. SLPs administered the WAB to get an observed AQ. Separately, patients underwent structural neuroimaging to collect lesion features (percent of voxels damaged within each region of the Johns Hopkins University atlas). With these features, we used Support Vector Regression (SVR) with Recursive Feature Elimination (RFE) and leave-one-out cross-validation to train a model on data from 70 patients and predict AQ for 1 left-out patient. Briefly, our RFE protocol trained the model to use only the top 10 most informative features for each split of the data, still keeping training and testing sets separate. We did this for each prompt individually, all prompts combined, lesion features only, and prompts+lesion features. We assessed model performance as the Pearson correlation between predicted and observed AQ (PredAQ; ObsAQ). Results: All model PredAQs were significantly correlated with ObsAQ (p < .05). Lowest performance was lesion-only (r=.61), highest was all discourse features combined (r=.83). Using discourse only, results for separate prompts were: BW (r=.78), PBJ (r=.7), and Cinderella (r=.74). Including lesion features with discourse did not significantly change model performance, except for Cinderella, which increased to r=.83. Inspecting feature weights revealed that the number of different grammatical types spoken was the most informative feature in multiple discourse-only models. Integrity of the superior longitudinal fasciculus was the most informative lesion feature. Discussion: This work represents an important step towards using discourse features to predict aphasia severity. While all prompts are informative, models trained on picture-sequence description output provide the highest prediction accuracy. Lesion-based models highlighted the importance of superior longitudinal fasciculus integrity for aphasia severity.
Topic Areas: Meaning: Discourse and Pragmatics, Disorders: Acquired