Clinical Report: Stroke Risk Assessment in ED Patients with Dizziness Using EMR-Based CDS

Overview

This study developed a clinical decision support system (CDS) leveraging electronic medical record (EMR) data and natural language processing (NLP) to predict stroke risk in emergency department (ED) patients presenting with acute dizziness. The model integrated structured clinical variables and unstructured physician notes, addressing limitations of current diagnostic tools and improving early stroke detection within seven days post-ED visit.

Background

Acute dizziness is a common ED complaint with a broad differential diagnosis ranging from benign vestibular disorders to life-threatening stroke. Stroke is frequently missed in dizzy patients due to subtle neurological signs and limitations of routine imaging and bedside tests. Existing clinical tools like ABCD2 and HINTS have variable accuracy and require specialized expertise, which may not be feasible in busy ED settings. Incorporating EMR data, including structured clinical variables and unstructured text, offers an opportunity to improve stroke risk stratification in this population.

Data Highlights

The study retrospectively analyzed ED visits from 2012 to 2021 at a tertiary hospital, including patients aged ≥20 years presenting with dizziness-related complaints. Stroke cases were defined as those diagnosed within seven days post-ED visit, while controls had no stroke diagnosis within one year. Features included demographics, vital signs, laboratory results, inflammatory markers, and physician notes processed via bag-of-words and BERT NLP methods. Missing data were imputed using multivariate imputation by chained equations (MICE), and continuous variables were normalized.

Key Findings

• Stroke is missed in up to 35% of ED visits for dizziness, highlighting diagnostic challenges.
• Routine neurological exams and CT scans have limited sensitivity for posterior circulation strokes.
• Existing clinical tools (HINTS, STANDING, TriAGe+) require specialized neurological assessment, limiting their ED applicability.
• Known stroke predictors in dizzy patients include age, male gender, diabetes, atrial fibrillation, prior cerebrovascular disease, recurrent vertigo, and elevated blood pressure.
• Integrating structured EMR data with unstructured physician notes via NLP enhances stroke risk prediction accuracy.
• Imputation and normalization techniques effectively handled missing and outlier data, improving model robustness.

Clinical Implications

Emergency physicians can leverage EMR-based clinical decision support tools incorporating both structured data and physician notes to better identify patients at high risk of stroke among those presenting with dizziness. This approach may reduce missed strokes and optimize resource utilization by guiding appropriate imaging and specialist consultation. Implementing such CDS systems can complement existing clinical assessments, especially in busy ED environments with variable staff expertise.

Conclusion

The study demonstrates that a machine learning model utilizing comprehensive EMR data and NLP can effectively predict stroke risk in ED patients with acute dizziness, addressing a critical gap in emergency stroke diagnosis. This CDS system holds promise for improving early stroke detection and patient outcomes in emergency care settings.

References

1. Edlow AG et al. 2017 -- Diagnosis and Management of Dizziness in the Emergency Department
2. Kattah JC et al. 2009 -- HINTS to Diagnose Stroke in the Acute Vestibular Syndrome
3. Kerber KA et al. 2015 -- Missed Stroke in Acute Vertigo and Dizziness
4. Newman-Toker DE et al. 2013 -- Diagnosing Stroke in the Emergency Department
5. Johnston SC et al. 2007 -- ABCD2 Score for Stroke Risk Stratification
6. Kattah JC et al. 2009 -- HINTS Exam Accuracy
7. Casani AP et al. 2016 -- STANDING Algorithm for Vertigo
8. Lee H et al. 2018 -- TriAGe+ Score for Central Vertigo
9. Ditmanson Research Database Documentation 2021