Abstract A134: Leveraging emergency medical services narrative text to understand contributors to missed stroke recognition in the field

Background: Emergency Medical Services (EMS) play a critical role in stroke systems by identifying stroke cases and facilitating rapid treatment. These benefits depend on accurate recognition, yet stroke nonrecognition remains frequent, and the mechanisms contributing to diagnostic failure are poorly understood. Methods: We conducted a retrospective cohort study of EMS-transported stroke cases using linked data from the Michigan Stroke Program Registry (MiSP) and the Michigan EMS Information System. Three EMS physicians manually reviewed EMS narratives and developed a dictionary of potential contributors to missed stroke recognition. After iterative refinement and blinded validation, we applied 65 narrative-based labels to all cases using regular expression techniques. We then fit a multi-group structural equation model (SEM) to examine associations between missed stroke recognition and four hypothesized latent constructs: distraction, classic hemispheric presentation, nonstroke symptoms, and patient participation limitations. Model development was guided by modification indices, residual diagnostics, and clinical plausibility. Results: Among 10,983 EMS-transported stroke cases (mean age 72; 51% female; 74% non-Hispanic white), 4,495 (41%) were not recognized as stroke by EMS. The final SEM demonstrated good fit (Comparative Fit Index = 0.914; RMSEA = 0.037, 90% CI: 0.035–0.038). Distracting situations were strongly associated with missed stroke recognition (β = 0.38, 95% CI: 0.27–0.48), primarily driven by nonstroke symptoms (loading = 0.75, 95% CI: 0.62–0.87). Classic hemispheric presentations were protective (β = –0.98, 95% CI: –1.04 to –0.93). Functional limitations had opposing effects: a direct protective association (β = –0.44, 95% CI: –0.60 to –0.29) and an indirect risk-enhancing effect via distraction (β = 1.23, 95% CI: 1.07–1.39), resulting in a non-significant net effect (β = 0.02, 95% CI: –0.04 to 0.08). Female sex (β = 0.02, 95% CI: 0.00–0.03) and nonwhite race (β = 0.05, 95% CI: 0.03–0.07) were independently associated with increased odds of missed stroke recognition. Conclusion: Distraction from atypical symptoms, functional complexity, and communication limitations contribute to EMS stroke recognition errors. These mechanisms interact with race and gender, underscoring the need for context-sensitive decision support to reduce disparities and improve prehospital stroke recognition.