Bringing doctors together

or

A free, private network for verified physicians

Leading voices

Dr. Connor O'Brien

Cardiac Critical Care · UCSF Medical Center

View

Dr. Rosy Thachil, MD, MBA

Cardiac Critical Care · Mount Sinai

View

Dr. Dorairaj Prabhakaran

General Cardiology

View

Dr. Pugazhendhi Vijayaraman

EP · Geisinger

View

Dr. Robert Schaller

EP · HUP

View

Dr. Trey Woods

General Cardiology · Emory Healthcare

View

Dr. Brahmajee Nallamothu

Interventional Cardiology · UMHS

View

Dr. Felix Yang

EP · NYU Langone

View

Dr. Jennifer Rymer

Interventional & Peripheral Cards · Duke University

View

Dr. Ami Bhatt

Adult Congenital Heart Disease · MGH

View
  • Discover your community

  • Stay current

  • A profile that represents you

  • Find your people

About RoonPrivacyTerms

    Overview

    Comparing Phenotypes for Acute and Long-Term Response to Atrial Fibrillation Ablation Using Machine Learning.

    Feb 10, 2025DOI: 10.1161/circep.124.012860Journal ArticleAHA

    Written by: Prasanth GanesanMaxime PedronRuibin FengAlbert J RogersBrototo DebHui Ju ChangSamuel Ruiperez-CampilloViren SrivastavaKelly A BrennanWayne R GilesTina BaykanerPaul CloptonPaul J WangUlrich SchottenDavid E KrummenSanjiv M Narayan

    Find these authors on Roon

    BD

    Dr. Brototo Deb

    Cardiac Critical Care · BWH

    PW

    Dr. Paul Wang

    EP · Stanford

    Introduction

    Atrial fibrillation (AF) is the most common arrhythmia affecting 30 million people worldwide, in whom it may cause hospitalization, heart failure, or stroke 1. It is increasingly shown that early rhythm control of AF reduces morbidity and mortality compared to ventricular rate slowing 2, and that rhythm control by ablation is more effective than medications in several patient groups 3,4. Nevertheless, the long-term success of ablation remains ∼50–70% despite several proposed additional lesion sets 5,6, and it is difficult to identify potential responders 7,8.We hypothesized that phenotyping may be improved by contrasting phenotypes between two relatively discordant endpoints to the same treatment in the same patients. In AF ablation, the primary endpoint is arrhythmia-freedom at 1 year. A distinct endpoint is acute termination of AF during the procedure. While acute termination is mechanistically plausible, its link to long-term success is weaker for AF than for atrial flutter or atrial tachycardia (AT) 9,10, or supraventricular tachycardias like AVNRT or WPW ablation 11,12. Unlike other arrhythmias, AF patients without any discernable acute impact of ablation may still be arrhythmia-free at 1 year or beyond 1. It is undefined whether the discordance between acute and chronic outcomes reflects demographic or lifestyle factors, differences in pathophysiology, or technical factors. The answers to this question have implications for intra-procedural assessment and may provide mechanistic insights.We hypothesize that different combinations of electrophysiological and clinical features separate phenotypes for long-term or acute response to AF ablation, and that probing these phenotypes by explainability analysis of ML may reveal mechanistic insights (Graphical abstract). We tested this hypothesis in the Stanford AF Ablation cohort (N=561) and an external AF cohort (N=77) of well-characterized patients, both undergoing a largely uniform ablation strategy to attenuate the contribution of procedural variability.

    Methods

    The study design is outlined in figure 1.The study was approved by the Stanford Institutional Review Committee and all subjects gave written informed consent. The Stanford AF ablation registry comprised N=726 consecutive patients referred for AF ablation for routine indications between 2014–2021. We excluded patients with incomplete long-term follow-up due to leaving the Stanford system (see Consort diagram in supplemental figure 1), leaving N=561 patients for this study. We split the data 70% (N=393) for training and 30% (N=168) as a hold-out set for testing, stratified to ensure that sets were balanced for age, sex, and endpoints. Details of stratification are provided in supplemental table 1. A subset of deidentified data used in this study is available from the corresponding author upon reasonable request.All patients were studied in the post-absorptive state, with antiarrhythmic medications discontinued >5 half-lives (>30 days for amiodarone). Under general anesthesia, catheters were placed in the right atrium (RA), CS, and left atrium (LA) via transseptal puncture. Electroanatomic mapping was performed using NavX (Abbott, IL) or Carto (Biosense-Webster, CA) that in all patients included global AF electrical data from 64-pole multipolar (basket) catheters placed in both atria (Abbott, IL). For patients who presented in sinus rhythm, AF was induced with rapid pacing from the CS. If atrial tachycardia (AT) was observed or induced, this was mapped and ablated first, followed by AF induction. Basket catheters offer a global mapping approach that continues to be used 13 and, despite suboptimal contact, Honarkbaksh, Schilling et al. reported ~75–80% atrial coverage with this design 14.N=293 presented in atrial fibrillation. N=22 patients presented in AT/flutter. N=246 patients presented in sinus rhythm. Mapping started after at least 5 minutes of induced sustained AF; in practice, mapping started 20–30 minutes after induction. Ablation was guided by clinical mapping systems (Carto, Biosense Webster, or NavX, Abbott). PVI was performed by guidelines, with wide-area circumferential radiofrequency ablation (WACA) or cryoablation (Medtronic, MN) of left and right pulmonary veins, and verification of entrance and exit block. Ablation strategy for each patient was driven by operator preference. Patients had (i) PVI/WACA (N=487) with additional mapping and ablation of localized sites identified as rotational or focal activity, or sites of complex activity; (ii) additional ablation such as roof or mitral lines (N=38); a minority of patients had isolated PVs and underwent ablation of localized rotational, focal or complex activity (N=36).Acute outcome was considered as success if AF terminated during ablation. To minimize the inclusion of chance events, terminations were included only if AF had been sustained for at least 30 minutes prior to that time. The need for cardioversion to terminate AF was considered a failure of the acute endpoint.To define long-term outcome, we applied a 3-month post-ablation blanking period in which repeat ablation was not permitted. Patients were evaluated in clinic for recurrence using external ambulatory ECGs quarterly for 1 year (mostly Zio, iRhythm Inc., California). This protocol was recommended, with 76.6% adherence (N=430). All patients were followed through our electronic medical system (Epic, Verona, WI), including records from outside hospitalizations or clinic visits using Care Everywhere. Long-term freedom from AF was defined as freedom from all atrial arrhythmias, i.e., AF/AT, defined as >30 seconds on any ECG monitoring.We exported the following electrical, imaging and clinical data.EGM: 1-minute recordings of EGMs from the LA (N=519) and N=333 (60% of dataset) had EGMs from both LA and RA. Electrograms were filtered at 0.05 to 500 Hz, notch filtered at 60Hz and exported from Bard (N=231) or Prucka (N=330) recording systems. We subsequently subtracted a mean QRS complex and filtered high frequency noise using a fourth order Butterworth bandpass filter (1 to 150Hz).12-lead ECGs obtained closest to the ablation procedure in sinus rhythm (N=234) or AF (N=327) were included. ECG features were annotated from electronic health records (EHR).Structural measurements from echocardiography were obtained from EHR reports.Clinical and demographic data: Variables including age, sex, AF duration, and lifestyle variables were extracted from EHR (full list in Supplemental Table 2). We obtained 51 clinical variables after excluding those with >30% missing values.Supplemental table 1 shows stratification performed for training and test splits.To identify features of AF contributing to long-term and acute outcomes, we extracted electrical, imaging, clinical and lifestyle features (72 total features).Electrical features: We first divided 64-channel EGMs from each atrium into non-overlapping 10-second segments to account for temporal variability. We computed the following 5 features: dominant frequency (DF) in Hz 15,16, cycle length (CL) in milliseconds 17, peak-to-peak voltage in millivolts 18, half-maximum width of DF peak (organization index or OI) 19, and similarity of EGM morphologies over time (maximum autocorrelation) 20. These features were averaged spatially (for each atrium) and across time-segments to obtain a feature-matrix of dimension 561×10. We included well-reported ECG features 21 of R-R interval, P-R interval, corrected QT interval, P-axis, QRS-axis and T-axis. This resulted in a feature-matrix with a dimension of 561×6.Imaging features: LVEF, LVH and LA size resulted in 561×3 feature-matrix.Clinical features have a dimension of 561×51, i.e., 51 values of feature per patient.Lifestyle Feature: History of smoking and alcohol consumption were added as binary variables (561×2 feature matrix).Features were averaged spatially (i.e. across 64 electrodes), and temporally (across 10 second segments) to obtain a single feature value per patient. The process was repeated for each feature. Missing data were imputed using substitution of mean for continuous variables and by addition of an “unknown” category for categorical variables 22.We used the 561×72 feature-matrix to train 6 ML models (figure 1). We used the widely popular grid search approach 23 on 5 fold cross validation of the training set to tune hyperparameters. This may eliminate bias 23 and overfitting that may occur with a fixed validation set. The test set was completely separate to avoid contamination. For each outcome, we selected the best performing model on the hold-out test set for explainability analysis. We calculated feature contribution using the Shapley method 24 on this model on the training set, then normalized the values with respect to the sum, and compared the top contributing features between the two outcomes. We then assessed the correlation of Shapley values between the two outcomes. For this purpose, we find the average Shapley across the training set for each feature resulting in 72 values which were compared between the outcomes.For generalizability, we tested our models on an independent population of N=77 patients from UCSD and VA San Diego Medical Centers. From an initial dataset of N=186 patients, we excluded patients without unipolar recordings (i.e. only bipolar recordings), and without both acute and 1-year outcome labels. The baseline characteristics of the external dataset are provided in supplementary table 4.This external registry had a data repository of which N=39 features overlapped with the Stanford registry: N=32 clinical features, 2 imaging features, and 5 electrical features from LA EGMs (see supplement table 5). We retrained our best ML models in the Stanford dataset using these 39 features, to ensure that the trained ML model could be tested directly on both cohorts. We calculated AUROCs and used Shapley to visualize top 10 features for each outcome.Continuous data are represented as mean ± standard deviation unless otherwise indicated. AUCs between the outcomes compared using Z tests and accuracy compared with McNemar’s test. Two sample t-test was used to assess p-values between two groups of continuous, normally distributed variables. Mann Whitney U test was used to assess p-values if the distribution is skewed. Categorical variables were compared using the chi-square test. Correlations between Shapley scores of outcomes were determined using Pearson correlation and that between binary outcomes were determined using the Phi coefficient (also known as Mathew’s Correlation Coefficient or MCC).

    Results

    Table 1 summarizes baseline characteristics. Patients with long-term freedom from AF were less likely to have prior failed AF ablation than those with AF recurrence (28.5 vs 41.5%, p=0.002). Conversely, patients with acute AF termination were more likely to have had prior AF ablation compared to those without acute termination (36.3 vs 28.3%, p=0.04). Supplementary table 3 shows demographics for patients only at first time ablation, split by long term and acute response.Patients with good term outcomes had lower age, left atrial size, and CHADS2VASc score compared to those with unfavorable outcomes, consistent with the literature 25. There was no difference in baseline characteristics such as CHADSVASC score (p=NS) nor long-term success in patients with or without the use of quarterly ambulatory ECGs (70.2% vs 67.2%, p=0.50, NS).We found that model predictions correlated poorly between acute and long-term outcomes in our training set (N=393). Figure 2A indicates the raw numbers of patients who fell into each group, and the groups showed a Phi coefficient of 0.08 indicating little to no relationship.The predictive performance of all 6 models are summarized in figure 3. The best model for acute outcome (Random Forests, an ensemble of decision trees) performed better than the best model or ensemble of six models for long-term outcome (logistic regression) (AUC 0.86 vs 0.67, p<0.001 McNemar test, p=0.0012 Z test; figure 3A), with higher predictive power for acute than long-term outcome regardless of the type of ML model (figure 3B).Explainability analysis of the best performing ML model revealed different predictors for long-term success (figure 4A) versus acute termination (figure 4B). In figures 4A and B, the x-axis represents a likelihood of success or failure of outcome and the colors represent the feature value, i.e. high or low value for continuous variables, yes or no for categorical variables.For acute AF termination, the top predictive features included patients with induced AF, paroxysmal AF, high EGM similarity of LA, longer RR interval and other features indicated in figure 4A. For long-term outcomes, the top features included smaller LA size, patients with induced AF, no history of smoking and other features indicated in figure 4B. Remaining features had low Shapley scores (<0.5). Notably, the need for AF induction was highly predictive for both acute termination and long-term success, independent of AF type.Predictors for long-term AF freedom poorly predict acute termination and vice versa (R2=0.13, figure 4C) and were largely discordant except for AF induction.Results were maintained when applying the trained ML model to patients receiving ablation for the first time (N=380; supplementary table 3) compared to those at redo procedures (fig.5).Overall, acute outcome was predicted primarily by electrical features in redo and first time ablation patients, which explained 70–80% of variance. Predictors of long-term outcome remained predominantly clinical and lifestyle factors, with some changes in the rank order of features (red arrows in Figure 5). Predictors of acute success were also similar between patients at first time and redo ablation (figure 5 C&D).We repeated ML analysis in patients with non-paroxysmal AF (N=338) compared to patients with paroxysmal AF. For long-term success, ninety percent of features were unchanged (figure 6 A&B) with some ranking changes. Notably, ‘induced AF’ was a greater contributor to long-term success in non-paroxysmal than paroxysmal AF patients. Similarly, for acute termination, 90% of features were similar with little change in rank order between paroxysmal and non-paroxysmal patients (figure 6 C &D).We found no significant difference (p=NS) between ablation strategies between patients with good and poor long term or acute outcomes (supplemental figure 2).Figure 7 shows the top 10 features contributing to the prediction of acute and long-term outcomes. AUC for predicting acute outcome was 0.81, similar to the internal dataset (AUC 0.83). AUC for long-term outcome was 0.64, similar to the internal dataset of 0.65. Similar features explained predictions in the external validation cohort as our primary cohort.

    Discussion

    In a large diverse cohort of patients undergoing AF ablation, distinct pathophysiological features predicted patients who responded acutely to ablation, versus those with arrhythmia-freedom at 1 year. Using a combination of reproducible ML models and novel approaches to explainability, we found that that arrhythmia-freedom was predicted modestly and mostly by clinical and lifestyle features, while acute termination was predicted accurately by electrical features. Notably, phenotypes for each outcome were poorly related. Remarkably, factors were consistent when comparing patients undergoing first time or repeat ablation, those classified as paroxysmal versus non-paroxysmal AF, and in an independent external registry.These data support the notion that physiological differences rather than technical factors explain the poor correlation between acute and long-term outcomes. It is notable that the need to induce AF for mapping linked both outcomes, independent of AF type. This further motivates strategies to restore sinus rhythm prior to ablation in patients with persistent AF, and suggest that AF burden could guide referrals to AF ablation. These data also add urgency to identify features of AF progression that better link acute and long-term outcomes.While several previous studies have predicted either acute or long-term outcomes from AF ablation, few have systematically compared predictors across clinical, lifestyle, electrophysiological, and structural domains. We applied competing classification and machine learning models, which is also rarely reported, to a richly characterized cohort of N=561 patients with 72 features, including the uncommon availability of global bi-atrial electrophysiologic mapping during AF in all patients.Long-term outcome in our study was predicted by LA size, confirming the majority of studies in the literature26. Importantly, the need for AF induction ranked second, and was one of the few features that also linked acute termination. A prior report identified sinus rhythm as a predictor of long-term success 27, but did not correlate this to acute outcomes. AF induction was more predictive than the clinical label of AF type, which is actually consistent with studies that clinical AF type poorly reflects AF burden on implanted ECG monitors 28.Our study supports previously reported relationships between long-term outcome and alcohol consumption and smoking29. The predictive values for long-term and acute outcomes in our study agrees with those reported by Lankveld, Schotten et al 30 in 93 patients of AUC = 0.61 and AUC = 0.70, respectively, who did not study intracardiac electrograms 31. Earlier studies found that ECG features predicted acute outcomes while AF duration predicted long-term success 26. Other studies of only long-term outcomes showed that clinical variables provide similar AUC of 0.72 8. A prior study from our group 22 in N=156 patients found that the top contributing features for long-term success from AF ablation was LVEF. A series of studies show that prediction of long-term success using CHADS2VASc score is lower at about 0.63 32.This work extends our previous work33,34 that spatial organization in AF electrograms in 1 cm2 regions (REACT index) predicted AF termination by ablation (AUC 0.72), and studies that ablation targeting localized drivers may improve outcomes35,36. Other studies report a series of intracardiac indices that predicted acute response 37 including dispersion of electrograms 38, and high dominant frequency 39.Prediction of long-term outcomes from AF ablation by clinical factors support the notion that disease progression may be a predominant driver of recurrence, beyond baseline features. Progression may reflect atrial myopathy, although this may be less explanatory of early recurrence 40. Acute termination may reflect closing of critical isthmuses, reduction of critical mass for reentry 41 or elimination of driver mechanisms 37–39. AF could then recur long term due to lesions that recovers, or mechanisms neither identified nor operative during that AF episode 42. These findings motivate a search for features of progression, or those that may link acute and chronic endpoints such as left atrial strain, autonomics or other electrical features. Our findings that the need for AF induction contributes to improve outcomes suggests either that patients with earlier disease (who present in sinus rhythm) or with more ‘controlled’ AF respond better. This motivates approaches to use AF burden to guide referral for early ablation 2,43 and to restore sinus rhythm prior to ablation. Novel indices are also needed to identify predictors of success in patients with already-isolated PVs undergoing redo ablation. These results could also form the basis for studies on novel machine-learned phenotypes.Our study did not include structural features from magnetic resonance imaging or geometric factors from computed tomography. Such studies are planned, including automated analysis of imaging metrics such as LA sphericity index for which we have developed machine learning tools 44. Although patients had paroxysmal and non-paroxysmal AF, first-time and redo-ablation, our sensitivity analyses show that predictive features were similar across sub analyses. Our population is large for a single center registry, and was concordant for an independent external registry, but could be further enlarged. Nevertheless, our single center design provided advantages of highly granular data, mapping of bi-atrial electrograms which may be difficult across multiple centers, and reduced operator variability. Although acute termination could in theory be a chance event, this may be less likely in the ~2/3rd of patients with non-paroxysmal AF and because we included AF termination only after long periods of sustained AF. Our ablation strategy were driven by operator preference, yet results from several strategies largely regress to those from PVI 45, and should not affect these results (see Supplemental figure 2).

    Conclusions

    In a large cohort of patients at AF ablation, machine learning revealed that long-term and acute response reflect distinct physiology. Long-term phenotypes were mostly, yet modestly predicted by clinical and lifestyle features, which may imply propensity for AF progression, while acute phenotypes were highly predicted by electrical features. The need for AF induction was the only feature to strongly predicted both outcomes. These results motivate studies to better characterize ‘early’ versus ‘progressive’ AF, and to restore sinus rhythm prior to ablation.
    BDDr. Brototo Deb

    Dr. Brototo Deb

    Cardiology · Cardiac Critical Care
    Brigham and Women's Hospital
    brototo