DOI: 10.1161/circ.148.suppl_1.14367 ISSN: 0009-7322

Abstract 14367: Multimodal Artificial Intelligence Algorithm to Distinguish Paroxysmal versus Persistent Atrial Fibrillation and Predict Outcomes After Ablation

Hannah Bernstein, Wesley Brooks, Padmini Sirish, Oliver Kreylos, Nipavan Chiamvimonvat, Alexandra McCann, Vladimir Filkov, Uma N Srivatsa
  • Physiology (medical)
  • Cardiology and Cardiovascular Medicine

Introduction: We aimed to develop effective multimodal machine learning (ML) techniques to predict: 1) duration of AF; 2) arrhythmia recurrence after ablation.

Methods: Clinical data, intracardiac EGMs and metabolomic samples were obtained from patients undergoing AF ablation. Patients were prospectively followed to assess for recurrence after a 90-day blanking period with monitors every three months and if symptomatic. A prototype model was implemented via projection onto latent structures (PLS) logistic regression (LR), which accounts for collinearity in the predictors by estimating a small number of latent vectors that are aligned with the binomial response. PLS LR was used to estimate models for distinguishing paroxysmal from persistent AF and predicting recurrence. The latent structures identified by PLS LR combine strength from measurements of multiple correlated features. We generated a reference distribution to assess whether the model is better than chance by permuting the responses while keeping the inputs fixed. Models were compared based on the Akaike Information Criterion.

Results: Three models were compared: EGM data (n=39), metabolomic data (n=29), multimodal approach on the combined two datasets (n=23). The multimodal model consistently produced a better fit despite smaller sample size. Only the model based on EGMs was significantly better than chance at distinguishing PAF from PeAF (p=0.03). The models based on EGMs and the combined data were significantly better than chance at predicting recurrence (p=0.02, p=0.04, respectively). Predictor EGM and metabolomics variables are shown in tables.

Conclusion: Multimodal ML algorithm is promising to effectively combine multidimensional data for clinical use. The combined model consistently produced a better fit compared to the others, indicating that multimodal approaches are promising. According to AIC, the fit was better when the datasets were combined.

More from our Archive