Predicting atrial fibrillation in primary care using machine learning.

Area of Science:

• Cardiology
• Medical Informatics
• Machine Learning

Background:

• Atrial fibrillation (AF) is the most common heart arrhythmia, often asymptomatic, leading to delayed diagnosis and serious complications.
• Existing risk-prediction models for AF often lack real-world applicability due to outdated data or limited scope.
• There is a critical need for implementable, contemporaneous risk models using routinely collected patient data for effective AF detection.

Purpose of the Study:

• To develop and evaluate novel statistical and machine learning models for predicting AF risk.
• To compare the performance of new models against established AF risk-prediction tools.
• To identify both known and novel patient factors associated with AF risk.

Main Methods:

• A retrospective cohort study analyzed 2,994,837 adults from January 2006 to December 2016 using the Clinical Practice Research Datalink.
• Evaluated various models including published (Framingham, ARIC, CHARGE-AF), machine learning (neural network, LASSO, random forests, SVM), and Cox regression.
• Models utilized baseline and time-updated patient information for risk prediction.

Main Results:

• Time-varying neural networks emerged as the optimal model, achieving an AUROC of 0.827, significantly outperforming the CHARGE-AF model (0.725).
• The optimal model demonstrated a lower number needed to screen (9 vs. 13) at 75% sensitivity compared to CHARGE-AF.
• Identified key predictors including age, prior cardiovascular disease, medication use, proximity of cardiovascular events, BMI changes, pulse pressure, and blood pressure measurement frequency.

Conclusions:

• The developed time-varying machine learning model offers superior predictive performance for AF compared to existing risk models.
• The model effectively incorporates both established and newly identified time-varying risk factors for AF.
• This approach holds promise for improving the early and efficient detection of undiagnosed atrial fibrillation.