In a groundbreaking development set to transform cardiology practice worldwide, researchers at Mount Sinai have unveiled an advanced artificial intelligence (AI) model that fundamentally reshapes how anticoagulation therapy is administered to patients suffering from atrial fibrillation (AF). This innovative Graph Neural Network (GNN)–based AI system leverages vast electronic health record datasets to provide highly individualized treatment recommendations, striving to optimize outcomes by balancing stroke prevention with the risk of major bleeding events. Presented recently in a “Late Breaking Science” session at the European Society of Cardiology Congress 2025, this study heralds a paradigm shift in personalized medicine for AF, a common cardiac arrhythmia that affects millions globally.
Atrial fibrillation disrupts the heart’s normal rhythm, causing the atria—the upper chambers—to quiver ineffectively. This quivering can lead to blood stasis, fostering the formation of clots that, if dislodged, may travel to the brain and cause a devastating ischemic stroke. Currently, the clinical standard involves administering anticoagulants, or blood thinners, to nearly all AF patients to mitigate the elevated stroke risk. However, anticoagulants carry their own inherent danger: increased risk of significant bleeding, sometimes resulting in life-threatening hemorrhagic complications. Physicians must juggle the dual risks of stroke and bleeding, often relying on population-derived risk scores that inadequately capture individual nuance.
The Mount Sinai AI model operates distinctly from these traditional approaches by utilizing the entirety of a patient’s comprehensive electronic health record rather than generic risk scoring systems. This includes a deep dive into millions of data points derived from clinical visits, diagnostic codes, laboratory results, physician notes, and other health parameters. The AI synthesizes this vast trove through graph neural networks—a type of machine learning algorithm particularly adept at modeling complex relational data—to generate a net-benefit treatment recommendation rooted in patient-specific probabilities of both stroke and bleeding outcomes. Unlike conventional methods that apply average risk across populations, this model estimates risk at the individual level, accounting for intricate and subtle clinical features unique to each patient’s history.
The training of this AI encompassed an unprecedented volume of data, with researchers utilizing electronic health records from 1.8 million patients, aggregating over 21 million clinical visits, 82 million physician notes, and a staggering 1.2 billion individual data points. This immense dataset empowered the model to learn intricate patterns predictive of stroke and hemorrhage risk, culminating in a robust algorithm capable of balancing these risks dynamically. Subsequent validation was conducted internally on nearly 39,000 AF patients within the Mount Sinai Health System and externally on over 12,800 patients from Stanford’s publicly available datasets, affirming the model’s accuracy and generalizability.
Intriguingly, the AI recommended against anticoagulation therapy for approximately half of the patients who would have otherwise been prescribed blood thinners based on current clinical guidelines. This reclassification suggests a critical opportunity to reduce unnecessary bleeding complications without compromising stroke prevention. Such a reevaluation has profound implications for global healthcare, promising not only more effective individualized care but also potential reductions in medication-related adverse events and associated healthcare expenditures.
Beyond the clinical utility, this AI system offers practical benefits in reducing the cognitive burden on clinicians. Traditionally, doctors must mentally weigh multiple population-level risk scores and engage in complex trade-offs when advising patients. The Mount Sinai model simplifies this process by breaking down risk probabilities for stroke and bleeding distinctly for each patient, enabling clearer communication and shared decision-making. This ability to translate complex data into transparent, individualized risk profiles represents a meaningful advance in patient-centered care.
Moreover, the model is designed to continually update its recommendations dynamically, incorporating new data from patients’ evolving health records in real time prior to clinical appointments. This adaptability ensures that treatment decisions reflect the latest clinical information and patient status, further enhancing the precision and relevance of anticoagulation management over time. Such dynamic updating is crucial in conditions like AF, where patient risk factors and comorbidities can fluctuate substantially.
Experts involved in developing the AI emphasize that the model’s success demonstrates the transformative power of applying modern AI techniques—particularly graph neural networks—in healthcare. By integrating billions of data points, the system overcomes the limitations of “one-size-fits-all” risk scoring systems, presenting a truly personalized medicine approach. The researchers envision this as a template for future clinical AI applications that move beyond population averages and toward individualized, data-driven decisions.
Clinical leaders herald this innovation as a potential watershed moment. Dr. Joshua Lampert, Director of Machine Learning at Mount Sinai’s Fuster Heart Hospital, remarks that this AI method represents a “profound modernization” in AF management. He highlights the model’s ability to complement clinician judgment by offloading computational complexities and providing clear, patient-specific guidance. This, according to Dr. Girish Nadkarni, Chair of the Windreich Department of Artificial Intelligence and Human Health, signals a “true paradigm shift” toward precision anticoagulation strategies that could revolutionize patient care and clinical workflows.
Furthermore, clinicians see the model’s implications extending far beyond immediate treatment decisions. By giving patients a transparent understanding of their personalized risks and expected benefits, the system empowers more informed discussions and enhances shared decision-making. This aligns perfectly with current healthcare goals emphasizing patient autonomy and individualized care strategies.
Despite this promising breakthrough, the researchers note that further clinical trials are necessary to verify the real-world impact of AI-guided anticoagulation recommendations. If forthcoming randomized studies confirm even a fraction of the model’s predictive power observed in retrospective analyses, it could translate into substantial improvements in clinical outcomes and quality of life for millions of AF patients worldwide. This would mark one of the most significant advances in anticoagulation therapy in decades.
Mount Sinai’s stature as a global leader in cardiology and advanced healthcare lends additional credence to the study. Their Fuster Heart Hospital ranks among the top cardiology centers worldwide, reflecting a long-standing tradition of pioneering heart research and clinical excellence. This AI-driven innovation aligns with their commitment to integrating cutting-edge technology with clinical practice to improve patient care and safety on a broad scale.
In conclusion, the Mount Sinai-developed Graph Neural Network-based AI model stands at the forefront of precision medicine, offering a novel, individualized strategy for anticoagulant decision-making in atrial fibrillation. By meticulously analyzing comprehensive patient data, the model promises to minimize devastating strokes and dangerous bleeding events more effectively than ever before. As this technology advances toward clinical adoption, it is poised to revolutionize both how physicians approach AF treatment and how patients participate in their care, ushering in a new era of personalized, data-driven cardiovascular medicine.
Subject of Research: Artificial Intelligence in Clinical Decision-Making for Atrial Fibrillation Treatment
Article Title: Graph Neural Network Automation of Anticoagulation Decision-Making
News Publication Date: September 1, 2025
Web References:
https://icahn.mssm.edu/about/artificial-intelligence
https://esc365.escardio.org/esc-congress/sessions/16815
Image Credits: Mount Sinai Health System
Keywords: Machine learning, Cardiology, Atrial Fibrillation, Anticoagulation, Stroke prevention, Personalized medicine
