Enhanced Prediction of Atrial Fibrillation in Patients With Ischemic Stroke Through Electronic Medical Records and...

Area of Science:

• Cardiology and Neurology
• Medical Informatics
• Machine Learning in Healthcare

Background:

• Atrial fibrillation (AF) is a significant, often undetected, risk factor for ischemic stroke, increasing mortality and disability.
• Current AF prediction models lack generalizability due to reliance on single-institution data and structured variables.
• Timely AF identification is crucial for preventive treatment to mitigate stroke risk.

Purpose of the Study:

• To develop a robust and generalizable AF risk prediction model for stroke patients using electronic medical records.
• To integrate structured clinical variables with features from unstructured clinical text for comprehensive patient health representation.
• To validate the model's stability and generalizability across multiple hospital datasets for routine clinical practice.

Main Methods:

• Analyzed data from 2 Taiwanese hospitals (LIH and CYCH) involving nearly 10,000 patients.
• Employed 5 feature engineering techniques for unstructured data and 6 resampling methods for data imbalance.
• Compared 9 classification algorithms, performing internal and external validation to assess model performance.

Main Results:

• The optimal model for LIH used only structured variables; the CYCH model excelled by integrating structured data with text-derived variables (TF-IDF).
• Key predictors consistently identified were E/A wave velocity ratio, left atrial size, and age.
• Ensemble learning-based machine learning models demonstrated superior performance after rigorous validation.

Conclusions:

• Developed predictive models for AF in ischemic stroke patients, showing improved performance with integrated data.
• Validated the generalizability and stability of machine learning models across different healthcare settings.
• Highlighted the potential of combining structured and unstructured data for enhanced AF risk prediction in clinical practice.