Automated Extraction of Stroke Severity From Unstructured Electronic Health Records Using Natural Language Processing

Area of Science:

• Neurology
• Medical Informatics
• Artificial Intelligence

Background:

• Electronic health records (EHRs) are valuable for stroke research but face challenges with data abstraction and missing information.
• Accurate stroke severity assessment is crucial for quality improvement and comparative effectiveness studies.
• Existing methods struggle to reliably extract or predict the National Institutes of Health Stroke Scale (NIHSS) score from EHRs.

Purpose of the Study:

• To develop and validate a natural language processing (NLP) model for extracting and predicting the NIHSS score from EHR clinical notes.
• To overcome limitations in EHR data abstraction for stroke severity assessment.
• To facilitate large-scale phenotyping of stroke severity.

Main Methods:

• A two-stage NLP model was developed using least absolute shrinkage and selection operator (LASSO) regression.
• Stage 1 used regular expressions to extract documented NIHSS scores.
• Stage 2 used LASSO to predict NIHSS scores from clinical documentation when scores were missing.
• The model was trained and tested on data from Massachusetts General Hospital and validated on the MIMIC database.

Main Results:

• The model achieved high accuracy in both the training/testing set (RMSE: 2.89, Spearman correlation: 0.92) and the validation set (RMSE: 2.20, Spearman correlation: 0.96).
• The NLP model demonstrated robust performance in extracting documented NIHSS scores and predicting missing scores.
• The study included over 4000 patients across two large datasets.

Conclusions:

• The developed NLP model enables automated, large-scale stroke severity phenotyping from EHRs.
• This approach can significantly enhance real-world quality improvement and comparative effectiveness research in stroke.
• The model provides a reliable method for overcoming data limitations in EHR-based stroke studies.