Machine learning-based identification of risk-factor signatures for undiagnosed atrial fibrillation in primary...

Area of Science:

• Cardiology and Health Informatics
• Utilizing machine learning for predictive modeling in healthcare

Background:

• Atrial fibrillation (AF) poses a significant risk for ischemic stroke, with over a third of cases undiagnosed.
• Effective AF screening is crucial, especially for stroke survivors, to mitigate associated risks.
• Identifying undiagnosed AF in primary care and post-stroke populations requires efficient screening strategies.

Purpose of the Study:

• To explore routinely available predictors of incident atrial fibrillation (AF) in primary care and post-stroke patients.
• To leverage machine learning methods for identifying high-risk populations for AF screening using German health claims data.
• To develop and validate predictive models for incident AF and AF post-stroke.

Main Methods:

• Analysis of German health claims data (InGef database, n = 1,476,391) for patients aged over 45.
• Application of machine learning techniques, including component-wise gradient boosting, to identify AF predictors from ICD codes.
• Internal and prospective validation of predictive models for incident AF and AF post-stroke.

Main Results:

• A model with 13 predictors achieved an AUC of 0.829 (internal validation) and 0.755 (prospective validation) for incident AF.
• Confirmed known AF risk factors (male sex, hypertension, heart failure, valvular heart disease, chronic kidney disease).
• Achieved an AUC of 0.67 (internal validation) and 0.766 (prospective validation) for post-stroke AF prediction.

Conclusions:

• Machine learning-selected, ICD-coded clinical variables effectively enhance the identification of patients at risk for newly diagnosed AF.
• This automated approach using readily available data can optimize AF screening in primary care and among stroke survivors.
• The findings support targeted screening to identify high-risk populations for atrial fibrillation.