Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Recovering NIHSS aphasia severity from acute-stroke clinical narratives
Poster Session F, Friday, October 2, 2:45 - 4:45 pm, Wangari Maathai
Roger Newman-Norlund1, John Absher2, Sigfus Kristinsson1; 1University of South Carolina, Department of Communication Sciences and Disorders, 2USC School of Medicine Greenville, Department of Psychiatry
Introduction. The NIHSS language item collapses heterogeneous post-stroke aphasia presentations into a coarse 0 to 3 ordinal, yet the clinical encounter that produces that score also produces hundreds of words of narrative description in the admission note. If aphasia severity can be recovered automatically from this narrative, retrospective EHR cohorts that lack coded NIHSS items become available for neurobiology-of-language research at cohort scale. We test this directly using expert-coded NIHSS-9 (best language) labels paired with the clinical text in which they were assessed. Methods. As part of the Stroke Outcome Optimization Project (SOOP), we assembled 1,227 admission notes (History & Physical and Neurology Consult) from 1,173 acute-stroke patients at a single academic medical center, after excluding 41 placeholder cross-reference notes that deferred entirely to a separate H&P or Consult. Each note carried an expert-coded NIHSS-9 score: 461 cases of no aphasia (level 0), 325 of mild to moderate (level 1), and 237 of severe (level 2). Notes were preprocessed to remove the templated NIHSS scoring grid that some institutions paste into note bodies, leaving only the narrative description. We trained one L2-regularized logistic regression per NIHSS-9 level on TF-IDF unigram and bigram features, evaluated with patient-grouped 5-fold cross-validation so that no patient appeared in both training and test folds. Results. Patient-grouped cross-validated ROC-AUC was 0.80 for no aphasia, 0.74 for mild-to-moderate, and 0.86 for severe (PR-AUC = 0.71, 0.49, 0.61; F_1 at the optimal threshold = 0.68, 0.56, 0.62). Top-weighted features for severe aphasia were clinically coherent and reflected established stroke phenomenology: left MCA, noxious, noxious stimuli, mute, not following, unable to assess. Features driving the no-aphasia class were correspondingly intact-language descriptors: oriented, speech without, without dysarthria, intact. The mild-to-moderate intermediate class was the hardest to recover, paralleling its known inter-rater difficulty in the underlying NIHSS scale. Conclusion. Acute-stroke clinical narratives contain enough language-phenotypic detail to recover NIHSS aphasia severity at moderate accuracy, with the resulting models keying on clinically meaningful descriptors of comprehension, expression, and lesion localization rather than surface artifacts. The approach provides a substrate for retrospective cohort-scale extraction of aphasia phenotypes from EHR text, supporting downstream neurobiology-of-language analyses including longitudinal recovery trajectories and lesion-symptom mapping in larger stroke cohorts.
Topic Areas: Disorders: Acquired, Computational Approaches