Abstract 12117: Machine Learning to Probe Variability in Patient Outcomes After Atrial Fibrillation Ablation - The Stanford AF Registry (start)

Introduction: A major challenge in improving atrial fibrillation (AF) ablation is the variability in outcome between patients, which is not captured by traditional biostatistics. Hypothesis: We hypothesized that machine learning (ML) can be used to probe phenotypes of variable response to AF ablation. Methods: We studied 632 patients with drug-refractory AF, enrolled prospectively for a uniform strategy of PVI plus map-guided AF ablation at Stanford Health. We applied unsupervised ML to 64 variables to identify features associated with freedom from AF or atrial tachycardia (AT) at 3 years. Results: Patients (N=632) were 65±10 years, 28.2% female, BMI 30.2±6 kg/m2, 59.7% with non-paroxysmal AF, and 70.0% at de novo ablation. At 1-year, freedom from AF and AF/AT were 77.5% (95% CI interval: 74.2%, 80.9%) and 70.1% at 1 year (66.5%, 73.8%). At 3 years, they were 55.5% (51.2%, 60.1%) and 48.6% (44.3%-53.3%) respectively, regardless of antiarrhythmic drugs (P=0.23). Unsupervised ML revealed 3 phenotypic clusters that crossed conventional AF labels (Fig1). An Early AF cluster had best outcomes (1Y/3Y: 79.6 /53.5%); a Non-remodeled AF cluster (69.8/52.0%) had most comorbidities, a substantial number of non-paroxysmal AF patients, yet better outcomes than a cluster characterized by atrial or ventricular Remodeling (66.4/42.3%) (p=0.028) (Table 1). Conclusions: In a large prospective AF registry, ML revealed clusters for outcomes, spanning conventional clinical labels and highlighting that patients with comorbidities may have good outcomes if they lack structural remodeling, and that electrical and structural remodeling may follow independent courses between patients.