This Week in Cardiovascular AI

Summarizing the recent literature on Cardiovascular AI.

Feb 23, 2025

Welcome to to This Week in Cardiovascular AI - a bi-monthly email newsletter for followers and subscribers that summarizes recent research in Cardiovascular AI that we find important and unique. For each edition we will summarize timely publications that are important in this space. This newsletter is a companion to CardioCyper, a long-form blog dedicated to the expanding intersection of Digital Health, Artificial Intelligence, Machine Learning and Healthcare, especially Cardiology and Electrophysiology.

The first article summarized below is the TAILORED-AF trial from Volta Medical. It investigated the use of their tech platform ‘AF-Xplorer’ and whether using this to guide ablation for persistent AF resulted in better outcomes than pulmonary vein isolation alone. This is an interesting trial and I will be doing a deep dive for the CardioCypher blog later in the week. Happy reading…

Share This week in Cardiovascular AI

AI-Guided Ablation Enhances Outcomes in Persistent Atrial Fibrillation: Results from the TAILORED-AF Trial

A recent Nature Medicine study by Deisenhofer et al. (2025) presents findings from the TAILORED-AF trial, a multicenter, randomized controlled study evaluating the effectiveness of AI-guided ablation for persistent and long-standing persistent atrial fibrillation (AF). The trial compared standard pulmonary vein isolation (PVI) with a tailored AI-guided ablation strategy targeting spatio-temporal electrogram dispersion in addition to PVI.

The results were promising: 88% of patients in the AI-guided group achieved freedom from AF at 12 months compared to 70% in the conventional PVI group. The AI-guided approach, while involving longer procedure times, resulted in significantly higher rates of arrhythmia elimination and improved outcomes, especially in patients with AF duration of 6 months or longer. Notably, while the tailored group showed more atrial tachycardia (AT) recurrences, these were more manageable with repeat ablation procedures.

The study highlights AI’s ability to objectively and consistently detect arrhythmogenic areas, allowing for more precise ablation strategies than traditional methods. Despite the longer procedural times, the AI-guided strategy demonstrated superior efficacy without increasing adverse events, offering a new standard of care for managing complex AF cases.

🔗 Read the full study: Nature Medicine

AI Revolutionizes Ambulatory ECG Analysis with Direct-to-Physician Reporting

A recent study in Nature Medicine by Johnson et al. (2025) showcases the transformative potential of DeepRhythmAI, an advanced artificial intelligence (AI) system designed for direct-to-physician reporting of ambulatory electrocardiograms (ECGs). Analyzing data from 14,606 patients over an average of 14 days of monitoring, the study compared the AI model's performance to that of 167 certified ECG technicians across a vast dataset of over 211,000 patient-days.

The AI model achieved a 98.6% sensitivity for detecting critical arrhythmias, including atrial fibrillation (AF), ventricular tachycardia (VT), asystole, and third-degree atrioventricular block (AVB), significantly outperforming technicians who achieved 80.3% sensitivity. Notably, the AI system reduced the risk of missed diagnoses by 14.1 times compared to technician-based analysis, with a false-negative rate of 3.2 per 1,000 patients versus 44.3 per 1,000 for human technicians.

While the AI system exhibited a modest increase in false-positive detections—12 per 1,000 patient-days compared to 5 for technicians—it maintained a negative predictive value exceeding 99.9%, making it highly reliable for detecting life-threatening arrhythmias. This high accuracy, combined with real-time analysis capability, offers the potential to reduce healthcare costs and enhance access to timely arrhythmia detection in patients requiring continuous ECG monitoring.

This research supports the implementation of AI-driven ECG analysis for routine clinical use, particularly in resource-constrained settings, offering a scalable and efficient solution for the growing demand for arrhythmia diagnostics.

🔗 Read the full study: Nature Medicine

SMARTBEATS-ALGO Study: AI-Driven Smartphone Monitoring for Atrial Fibrillation Detection

A new study published in Europace by Fernstad et al. (2025) highlights the effectiveness of a machine learning (ML) algorithm for diagnosing atrial fibrillation (AF) and atrial flutter (AFL) using smartphone-based photoplethysmography (PPG). In the SMARTBEATS-ALGO trial, researchers validated the algorithm’s performance in an unsupervised, real-world ambulatory setting among patients undergoing cardioversion for AF or AFL.

The study enrolled 460 patients, generating over 34,000 simultaneous PPG and ECG recordings using a smartphone PPG application (CORAI Heart Monitor) alongside a single-lead ECG device (KardiaMobile). The ML algorithm, based on support vector machines (SVMs), demonstrated outstanding diagnostic accuracy for detecting AF, achieving a 99.7% sensitivity, specificity, and accuracy. When AFL was included, the diagnostic accuracy remained high at 99.2%.

The study also showcased the potential of real-time feedback mechanisms within the PPG application, which significantly improved signal quality and diagnostic performance. Importantly, the algorithm effectively reduced the need for manual ECG reviews, making this technology particularly promising for remote patient monitoring and post-stroke AF screening in elderly populations.

This research highlights the growing role of AI-driven smartphone diagnostics in enhancing accessibility to non-invasive arrhythmia monitoring, potentially transforming how clinicians manage atrial arrhythmias in outpatient and telehealth settings.

🔗 Read the full study: Europace

AI-Powered Patient Selection for Left Atrial Appendage Occlusion in Atrial Fibrillation

A recent study in JACC: Clinical Electrophysiology by Ngufor et al. (2025) applies a novel causal machine learning (ML) framework to optimize patient selection for left atrial appendage occlusion (LAAO) in patients with atrial fibrillation (AF). Using data from over 744,000 AF patients treated with either LAAO or direct oral anticoagulants (DOACs), the study aimed to identify which patients would benefit most from LAAO compared to lifelong anticoagulation therapy.

In a matched cohort of 28,930 patients, LAAO showed no significant difference in adverse outcomes compared to DOACs at 1 year but demonstrated a 2.9% lower risk for a composite of ischemic stroke, systemic embolism, major bleeding, and all-cause mortality by the end of 2 years. Notably, 30.1% of patients were classified as likely to benefit from LAAO by the end of the second year, while 69.7% showed no clear advantage, and 1.4% were potentially at risk of harm.

The ML model, using a causal forest algorithm, identified key factors influencing LAAO outcomes, including advanced age, comorbid conditions (e.g., dementia, chronic obstructive pulmonary disease, and heart failure), and previous stroke events. This approach outperformed traditional statistical models in predicting heterogeneous treatment effects (HTEs), offering a personalized medicine tool for clinicians to better identify AF patients who may gain the most from LAAO.

This research underscores the potential of AI-driven decision support tools in refining patient selection for LAAO and improving clinical outcomes in AF management, particularly for those who face challenges adhering to lifelong anticoagulation.

🔗 Read the full study: JACC: Clinical Electrophysiology

Machine Learning Identifies Distinct Phenotypes for Acute and Long-Term Response to Atrial Fibrillation Ablation

A recent study in Circulation: Arrhythmia and Electrophysiology by Ganesan et al. (2025) applied machine learning (ML) to distinguish phenotypes associated with acute and long-term outcomes of atrial fibrillation (AF) ablation. Analyzing data from 561 patients in the Stanford AF Ablation Registry, the researchers compared six ML models to predict outcomes using 72 multimodal features spanning electrogram characteristics, ECG, cardiac structure, lifestyle, and clinical variables.

The study found that acute and long-term responses to AF ablation are driven by distinct physiological factors. The best-performing model for predicting acute AF termination (Random Forest) achieved an AUC of 0.86, while predictions for long-term arrhythmia-free survival (Logistic Regression) showed a lower AUC of 0.67. Electrical features, such as electrogram similarity and cycle length, were strong predictors of acute success, whereas long-term outcomes were influenced by clinical and lifestyle factors, including left atrial size, smoking status, and prior ablation history.

Interestingly, the need for AF induction during mapping was the only feature predictive of both acute termination and long-term success. These findings were validated using an external cohort of 77 patients, where the model achieved similar performance (AUCs of 0.81 for acute and 0.64 for long-term outcomes).

The study underscores the importance of tailoring patient selection and treatment strategies for AF ablation based on distinct predictors of short- and long-term success, emphasizing the potential of AI-driven personalization in arrhythmia management.

🔗 Read the full study: Circulation: Arrhythmia and Electrophysiology

Predicting Atrial Fibrillation After Stroke: Insights from a Systematic Review and Meta-Analysis

A new systematic review and meta-analysis published in Heart Rhythm by Helbitz et al. (2025) evaluates the performance of multivariable prediction models for detecting atrial fibrillation (AF) in patients following a stroke. The study reviewed 75 research papers covering 58 distinct prediction models and aimed to identify tools capable of stratifying stroke patients by their risk of developing AF, potentially guiding decisions for extended cardiac monitoring.

Three models demonstrated excellent discrimination in predicting post-stroke AF:

SAFE (C-statistic: 0.856, 95% CI: 0.796–0.916)
SURF (C-statistic: 0.815, 95% CI: 0.728–0.893)
iPAB (C-statistic: 0.888, 95% CI: 0.824–0.957)

However, when excluding studies with a high risk of bias or small sample sizes, only the SAFE model maintained excellent performance. Models such as CHASE-LESS and RE-CHARGE AF also showed adequate discrimination but fell short of the top-tier models in robustness and external validation.

Despite these promising results, the review highlights significant limitations, including a high risk of bias in 66% of the included studies and a lack of clinical impact assessments. No models have yet been prospectively validated in real-world clinical settings, leaving their practical utility in question.

The study concludes that while several models show potential for guiding extended monitoring in stroke survivors, prospective validation studies are essential before integrating these tools into clinical guidelines. This review underscores the need for more robust research to optimize stroke care and prevent secondary events linked to undetected AF.

🔗 Read the full study: Heart Rhythm

Machine Learning Identifies Distinct Phenotypes of Persistent Atrial Fibrillation: Insights from the DECAAF II Trial

A recent study published in the Journal of Cardiovascular Electrophysiology by Noujaim et al. (2025) applies machine learning (ML) to uncover distinct phenotypes of patients with persistent atrial fibrillation (AF) undergoing catheter ablation (CA). Utilizing data from the DECAAF II trial, which enrolled 815 patients across 44 centers, the researchers aimed to enhance patient selection for personalized treatment strategies based on post-ablation prognosis.

Using a Gradient Boosting Method (GBM) and k-medoids cluster analysis, the study identified two distinct patient clusters based on key variables: left atrial (LA) volume, body mass index (BMI), baseline fibrosis, and age. Patients in Cluster 1—characterized by older age, higher BMI, greater LA volume, and increased fibrosis—had significantly worse outcomes. Their arrhythmia recurrence rate was 51.7%, compared to 35.0% in Cluster 2. The survival analysis also revealed a 71% increased hazard of recurrence (HR = 1.71, p < 0.0001) for patients in the high-risk cluster. These findings were validated in an independent cohort from the trial.

The study underscores the heterogeneity of persistent AF and suggests that high-risk phenotypes could benefit from closer monitoring or more aggressive ablation strategies. By integrating machine learning into clinical decision-making, this approach offers a pathway to personalized ablation strategies, improving patient outcomes and procedural efficiency.

🔗 Read the full study: Journal of Cardiovascular Electrophysiology

The FDA’s Perspective on Digital Health Technologies for Cardiometabolic Disease and Diabetes

A recent Journal of the American College of Cardiology (JACC) publication by Warraich et al. (2025) explores the transformative potential of digital health technologies (DHTs) in managing cardiometabolic diseases, including diabetes, cardiovascular disease, obesity, and metabolic dysfunction-associated steatohepatitis. These technologies—ranging from continuous glucose monitors (CGMs) to smartwatches with ECG functionality—offer innovative ways to monitor health, support behavior change, and enhance disease prevention.

The paper highlights key challenges limiting DHT adoption:

Lack of evidence connecting DHT usage to improved long-term clinical outcomes, particularly for patients who need intervention the most.
Integration barriers within clinical workflows, as many DHTs fail to sync effectively with existing electronic health record (EHR) systems.
Limited patient access due to socioeconomic disparities, with underrepresented populations facing technological and financial barriers to adoption.

The FDA emphasizes the need for stronger clinical validation and real-world evidence to justify broader insurance coverage and integration into clinical guidelines. For instance, while CGMs are widely validated for type 1 diabetes, their impact on outcomes like weight, lipids, and blood pressure in type 2 diabetes remains unclear.

Additionally, the paper stresses the importance of developing AI-powered platforms capable of integrating disparate data from wearables and medical devices to offer personalized insights for both clinicians and patients.

To bridge the gap between innovation and adoption, the FDA recommends:

Designing inclusive studies that focus on high-risk, underserved populations.
Promoting interoperability between DHTs and existing healthcare data systems (e.g., using standards like HL7).
Involving clinicians and patients early in the development process to ensure user-friendly interfaces and clinical relevance.

This comprehensive perspective underscores the critical role of collaboration between regulators, clinicians, and technology developers in ensuring that digital health innovations reach their full potential in improving patient outcomes.

🔗 Read the full study: Journal of the American College of Cardiology

Machine Learning-Derived Left Atrial Wall Thickness Predicts Atrial Fibrillation Recurrence After Pulmonary Vein Isolation

A recent study published in the Journal of Cardiovascular Electrophysiology by Gomes et al. (2024) highlights the predictive value of left atrial wall thickness (LAWT), measured using machine learning (ML), for determining the likelihood of atrial fibrillation (AF) recurrence following pulmonary vein isolation (PVI). This single-center study analyzed 439 patients with symptomatic AF who underwent pre-ablation cardiac CT scans and follow-up over a median of 5.8 years.

The study found that higher LAWT was an independent predictor of AF recurrence post-ablation. Patients with thicker atrial walls were more likely to be older, male, and have more cardiovascular risk factors. Notably, LAWT remained a significant predictor even after adjusting for clinical variables such as left atrial volume index (LAVI), body mass index (BMI), and chronic kidney disease (CKD). The recurrence rates increased across LAWT terciles, with annual rates of 11%, 15%, and 21% corresponding to progressively thicker walls.

A subgroup analysis revealed a moderate correlation between increased LAWT and left atrial fibrosis, as measured by late-gadolinium enhancement on cardiac MRI. Importantly, regional analysis showed that posterior and anterior wall thicknesses were the strongest predictors of AF recurrence.

The findings suggest that LAWT, easily measured through AI-driven software, could serve as a valuable tool for risk stratification and personalizing ablation strategies. Patients with increased LAWT might benefit from tailored treatment plans, potentially improving long-term outcomes and reducing the likelihood of arrhythmia recurrence.