Identification of neurological text markers associated with risk of stroke.

Background: Delayed or missed stroke diagnosis is associated with poor outcomes. We utilized natural language processing of notes from non-neurological emergency department (ED) encounters to identify text phrases indicating stroke presentations that are associated with stroke hospitalization 30 days after ED discharge.

Methods: We conducted a retrospective analysis of stroke (case) and gastroenteritis (matched-control) patients at two academic medical centers who had an ED encounter 30 days before index admission diagnosis. Medical concepts were extracted from the ED encounter notes. Statistical analysis was used to detect neurological text markers indicating stroke signs and symptoms using data from one hospital (discovery cohort) and validated in the second (validation cohort). We further compared the coefficients and the predictive performance of an elastic net model of both cohorts.

Results: We detected 58 medical concepts with a statistically significant positive association with stroke cases in the discovery cohort of 987 patients (51 % stroke). Expert review was used to combine these medical concepts into 11 text markers indicative of stroke presentations (e.g., coordination, language). Markers demonstrated external validity in terms of positive association when analyzed in the validation cohort of 433 patients (24 % stroke). Elastic net models derived at each center demonstrated equivalence in coefficient magnitudes and predictive performance, demonstrating generalizability.

Conclusion: We detected and validated neurologic text markers characteristic of stroke signs and symptoms at an ED encounter 30 days before the stroke diagnosis. The presence of these markers could be used to prompt additional neurologic evaluation to prevent delayed stroke diagnosis.