Assessing stroke severity using electronic health record data: a machine learning approach.

Area of Science:

• Neurology
• Health Informatics
• Machine Learning

Background:

• Stroke severity is a key predictor of patient outcomes.
• National Institutes of Health Stroke Scale (NIHSS) scores are standard measures of stroke severity.
• Structured real-world evidence databases often lack NIHSS scores due to free-text reporting in physician notes.

Purpose of the Study:

• To develop and evaluate machine learning models for imputing NIHSS scores.
• To utilize multi-institution electronic health record (EHR) data for stroke severity assessment.
• To enable the incorporation of stroke severity into studies using administrative and EHR databases.

Main Methods:

• Extracted NIHSS scores from physician notes using natural language processing (NLP).
• Trained machine learning models on a cohort of 7149 patients with ischemic stroke, hemorrhagic stroke, or transient ischemic attack.
• Evaluated model performance using a holdout set of 1033 patients, with a random forest model showing optimal performance.

Main Results:

• Identified key EHR factors for stroke severity: mortality, hospital stay length, dysphagia, hemiplegia, and discharge destination.
• Achieved an R² of 0.57, an R of 0.76, and a root-mean-squared error of 4.5 when comparing imputed to NLP-extracted NIHSS scores.
• Demonstrated the feasibility of using machine learning to create stroke severity proxies from EHR data.

Conclusions:

• Machine learning models can effectively impute stroke severity (NIHSS scores) from EHR data.
• This method allows for the inclusion of stroke severity in large-scale studies using administrative and EHR databases.
• Imputed stroke severity enhances the analysis of patient outcomes and facilitates broader research in stroke epidemiology.