Simulating atrial fibrillation patterns and screening strategies using virtual patient-level models

Abstract Introduction Atrial fibrillation (AF) is the most common cardiac arrhythmia and is associated with increased morbidity and mortality. However, early detection and treatment of AF are hampered by highly variable occurrence and duration of AF episodes, heterogenous patterns of AF progression, and poor symptom-rhythm correlation present in many patients. Purpose To develop a computational patient-level model for simulating individual AF episodes, AF progression, and associated clinical outcomes, and to compare the effectiveness of systematic or symptom-based 1-minute electrocardiography (ECG) or 14-day Holter screening using this model. Methods We developed a Markov model with 7 different rhythm and/or clinical states (sinus rhythm (SR), asymptomatic and symptomatic AF, each with or without previous stroke, and death) and age/sex-dependent transition probabilities estimated based on clinical data (Figure 1A). Additionally, transitions between SR and AF states were modulated by AF-related remodeling, which itself was controlled by the time in AF and age. Results The virtual patient-level model enabled the simulation of changes in rhythm status and clinical outcomes during an entire lifetime with minute-level resolution (Figure 1B). Age/sex-specific AF prevalence / incidence, the proportion of symptomatic episodes and mortality in 4000 virtual patients were in good agreement with clinical data (Figure 1C). We employed the perfect knowledge about rhythm status at any moment in time provided by the model to evaluate systematic and symptom-based 1-minute ECG and 14-day Holter screening for AF in this population. In total, 2372 virtual patients >65 years developed AF. Systematic yearly 1-minute ECG recordings only identified 58 (2.4%) patients. In contrast, 14-day Holter-monitoring detected 18% of AF patients over 65 with a median of 532 days earlier than their clinical AF diagnosis, and identified 11.8% of AF patients who were either completely asymptomatic or only had symptomatic episodes <3 hours. Short-term (2-week) screening after a first symptom had a very low yield, which increased with increasing screening frequency and duration of follow-up (Figure 2). The majority of the remaining AF patients could be diagnosed when their symptoms lasted ≥3 hours, providing time for a traditional clinical diagnosis. Conclusion Our novel patient-level AF model can simulate individual AF patterns and clinical outcomes such as stroke and death in large cohorts of virtual patients. The model highlights the variability of AF episodes that makes effective early detection of asymptomatic AF with short-term monitoring challenging, and provides a tool for future studies into AF progression, screening and prevention of AF-related stroke and mortality.Figure 1Figure 2