Aim: The main objective of atrial fibrillation (AF) screening is to detect undiagnosed subjects, but collected data may also be used to identify participants at higher risk to develop AF in time, increasing the benefits and cost-efficiency of screening. This study aims at finding the optimal set of features that can be derived from screening data to perform long-term risk prediction of new-onset AF in an observation period of six years.
Methods: Based on the evidence of the prognostic implications of supraventricular arrhythmias (SVA) in AF development, two types of SVA-based features were designed to either inform about the presence or describe the characteristics (prematurity degree, burden and frequency) of SVAs detected in 30s single-lead ECG recordings. Different combinations of SVA-based features, median QRS complex duration, and clinical parameters were combined into ten investigated feature sets. Using data from 11104 participants of the Strokestop AF screening studies, risk prediction models are trained to perform multiclass classification of the participants into four risk groups using these feature sets, in order to evaluate the information embedded in each set by evaluating the respective model's performances. Performance of the models is assessed with risk group-specific AUROC and AUPRC, a task-specific cost function, and odds ratios.
Results: The risk stratification model trained on SVA characteristics and clinical parameters identifies a high risk (HR) group with 4.2 times increased odds of developing AF within two years after screening when compared to the lowest risk group, reaching an HR-specific AUROC of 0.67 and AUPRC of 0.06 (0.03 baseline). In addition, this model provides a statistically significant differentiation between subjects that develop AF at different times in the observation period.
Conclusion: Among the tested feature sets, SVA characteristics and clinical parameters constitute the optimal feature set for long-term risk prediction of new-onset AF from screening data.