Aug. 16, 2025
Researchers in Mayo Clinic Robert D. and Patricia E. Kern Center for the Science of Health Care Delivery in collaboration with Cardiovascular Medicine developed a new artificial intelligence (AI) algorithm to identify patients likely to benefit from transcatheter left atrial appendage occlusion (LAAO) versus direct oral anticoagulants (DOACs). This algorithm can help patients with atrial fibrillation (AFib) who need to be referred to a subspecialty for consultation. The findings were published in JACC: Clinical Electrophysiology.
Personalized stroke prevention
Worldwide, the estimated 33.5 million people with AFib face a fivefold increased risk of stroke. For stroke prevention in patients with AFib, lifelong oral anticoagulation is advised and clinical guidelines recommend DOACs for most patients. But oral anticoagulation increases the risk of bleeding. It also can be a challenge for patients to follow ongoing drug therapy, and many patients are undertreated.
"This research addresses a critical gap in personalized stroke prevention for patients with atrial fibrillation. While both LAAO and NOACs are used to reduce stroke risk, selecting the optimal therapy for individual patients remains a challenge," says Xiaoxi Yao, Ph.D., Endowed Scientific Director of Pragmatic Trials and Evaluation, professor of Health Services Research at Mayo Clinic Robert D. and Patricia E. Kern Center for the Science of Health Care Delivery, and senior author of the study. "This study introduces a novel causal machine learning framework that estimates heterogeneous treatment effects (HTEs), helping to identify which patients are more likely to benefit from LAAO over NOACs. This approach supports more precise, data-driven clinical decision-making and could reduce both undertreatment and overtreatment, ultimately improving patient outcomes and safety."
Which patients for LAAO?
LAAO is an alternative to lifelong anticoagulation. LAAO has been increasingly used as a stroke prevention strategy for patients with AFib. But which patients are good candidates? Selecting the best candidates for LAAO remains unclear in everyday clinical practice.
"The idea emerged from a critical clinical challenge of identifying which patients would truly benefit from LAAO. Traditional subgroup analyses and regression models often fall short — they either lack power, rely on oversimplified assumptions or fail to capture complex interactions among patient characteristics. To overcome these limitations, we turned to causal machine learning, specifically causal forest (CF) model to create a data-driven tool that could support more-personalized and equitable decision-making in stroke prevention, says Peter A. Noseworthy, M.D., M.B.A., a cardiac electrophysiologist, cardiologist and chair of Heart Rhythm Services at Mayo Clinic in Rochester, Minnesota, and one of the study's authors.
Several factors contribute to the difficulty in selecting the best candidates for LAAO:
Lack of individualized evidence. Most clinical trials and observational studies provide population-level recommendations, which may not apply to an individual patient.
Complex risk profiles. Patients with AFib often have multiple comorbidities such as dementia, lung disease and prior bleeding that influence both the risks and benefits of LAAO versus NOACs.
Limitations of traditional methods. One-variable-at-a-time subgroup analyses can miss important and complex interactions between multiple patient-level characteristics.
"This complexity has made it difficult to confidently identify candidates for LAAO referral," says Dr. Noseworthy.
Innovative cures
The algorithm can help in selecting patients to refer for further examination and possible discussion of LAAO. "Its ability to rank patients by benefit potential offers a new tool for triaging referrals, which could be transformative in systems with limited subspecialty access," says Dr. Yao.
For this study, the population included 744,190 adult patients treated with LAAO or DOAC between March 13, 2015, and December 31, 2019. Forty percent of the patients were female. The mean patient age was 76.8 years old.
Key findings show that 30.1% of patients were identified as likely to benefit from LAAO, while 69.7% were neutral and 1.4% potentially harmed. The machine learning model identified that older adults with higher comorbidity burdens were more likely to benefit from LAAO in comparison to DOACs.
"This innovation exemplifies how we seek to use existing data to make better decisions at the bedside," says Dr. Noseworthy. "By integrating causal machine learning into clinical workflows, cardiologists can tailor treatment decisions rather than rely on generalized guidelines. This research can enhance shared decision-making ;by providing patients with personalized risk-benefit insight."
Next steps
Looking ahead, researchers will test the algorithm in prospective studies or pragmatic trials to confirm its utility in real-world settings. They plan to refine the model with additional multimodality data and explore its applicability to other populations.
"Ultimately, this research bridges the gap between big data and bedside care," says Dr. Noseworthy. "It offers a scalable, explainable and clinically actionable solution that could be embedded in electronic health records systems to improve outcomes for patients with AFib nationwide."
For more information
Ngufor C, et al. Causal machine learning for left atrial appendage occlusion in patients with atrial fibrillation. JACC: Clinical Electrophysiology. 2025;11:977.
Refer a patient to Mayo Clinic.