For children born with tetralogy of Fallot, surgical repair marks not an end but the beginning of a lifetime of vigilance — a vigil that has long depended on expensive, inaccessible cardiac MRI scans. Researchers at Mount Sinai have now trained an artificial intelligence to read meaning into the humble electrocardiogram, finding within its familiar lines a signal of which hearts are quietly changing in ways that demand attention. The work, validated across five North American hospitals, does not seek to replace the gold standard of imaging but to make the long road of lifelong monitoring more
AI Tool Uses ECG to Predict Heart Changes in Tetralogy of Fallot Patients
Patients with repaired tetralogy of Fallot require lifelong monitoring; improved access to risk stratification could enhance long-term outcomes for this population.
A tool to make lifelong heart monitoring more accessible and efficient
Mark
So the AI reads an ECG and tells you whether someone needs an MRI. How accurate is it?
Mimi
It can estimate the risk of ventricular remodeling from the ECG patterns. But accuracy varied across the five hospitals where they tested it. That's the key finding—it works, but not uniformly.
Luke
What does "varied" mean in numbers? The source doesn't give sensitivity or specificity figures. We know it performed differently by site, but we don't know by how much.
Mark
Why would it perform differently at different hospitals?
Mimi
Different patient populations, different equipment, different ways of doing the scan or the ECG. The AI learned from one set of data, and when you move it to a new place, the real world is messier.
Luke
Right. And that's why they're saying local validation is critical before clinical use. It's not ready to roll out everywhere tomorrow.
Mark
So what's the actual benefit if you still need MRI to confirm?
Mimi
You don't do MRI on everyone. You do it on the high-risk patients the AI flags. That saves time, money, and imaging capacity for the people who need it most.
Luke
Assuming the AI's risk stratification is reliable at your hospital. Which brings us back to the validation question.
Mark
When do they think this will actually be in clinics?
Mimi
They're planning prospective trials next and refinement for younger patients. Long-term goal is integration into routine care, but that's years away.
Luke
And they're being honest that this is not a replacement for MRI. It's a gatekeeper. That's important to say clearly.
Il Polso
- Thousands of patients with repaired tetralogy of Fallot require lifelong cardiac surveillance, yet many miss critical MRI scans due to cost, access, and system bottlenecks.
- An AI model trained on routine ECG data can now flag which patients are at highest risk for ventricular remodeling — the silent heart changes that typically only MRI can detect.
- Validation across five hospitals revealed uneven performance by site, exposing a fundamental tension in clinical AI: a tool that works in one setting may quietly fail in another.
- Researchers are not positioning the AI as a replacement for cardiac MRI but as a triage mechanism — a smarter filter to direct scarce imaging resources toward those who need them most.
- Prospective trials and refinements for younger patients are planned, with the long-term ambition of weaving this tool into standard congenital heart disease follow-up care worldwide.
For children born with tetralogy of Fallot, surgical repair marks not an end but the beginning of a lifetime of vigilance — a vigil that has long depended on expensive, inaccessible cardiac MRI scans. Researchers at Mount Sinai have now trained an artificial intelligence to read meaning into the humble electrocardiogram, finding within its familiar lines a signal of which hearts are quietly changing in ways that demand attention. The work, validated across five North American hospitals, does not seek to replace the gold standard of imaging but to make the long road of lifelong monitoring more navigable — ensuring that the right scan reaches the right patient before the window of intervention closes.
A child born with tetralogy of Fallot undergoes surgery to repair a structurally compromised heart — but that surgery is a beginning, not a conclusion. What follows is a lifetime of monitoring, anchored by cardiac MRI scans designed to catch the slow drift of ventricular remodeling before it becomes irreversible. The problem is that MRI is expensive, time-consuming, and unevenly accessible. Many patients simply do not get the scans they need.
Researchers at Mount Sinai Kravis Children's Heart Center, backed by the National Institutes of Health, asked whether something already present in every clinic — the standard electrocardiogram — could carry more information than it appeared to. They built an AI model trained on paired ECG and MRI data, teaching it to estimate a patient's risk of ventricular remodeling from a test that takes minutes and costs a fraction of an MRI. The findings were published in the European Heart Journal: Digital Health.
Testing the model across five additional North American hospitals produced results that were promising but instructive. The AI demonstrated real predictive value, offering clinicians a way to prioritize which patients needed urgent imaging and which could safely wait. Yet performance varied meaningfully from site to site — a reminder that AI tools carry the biases and particularities of the environments that shaped them, and must be validated locally before they can be trusted clinically.
Lead author Son Duong described the work as an effort to unlock hidden value from a test that already exists everywhere. Co-senior author Girish Nadkarni, Mount Sinai's chief AI officer, framed the site-to-site variation not as failure but as a necessary lesson about the discipline required before AI enters routine care. The goal was never to retire cardiac MRI — it remains the definitive tool — but to build a smarter triage layer around it.
The team now plans prospective trials and model refinements aimed at younger patients, with a long-term vision of integrating the tool into standard congenital heart disease follow-up. If realized, the benefit would not come from replacing existing care but from reorganizing it — making sure that across a lifetime of monitoring, the right imaging finds the right patient at the right moment.
A child born with tetralogy of Fallot—a congenital heart defect affecting the structure of the heart—undergoes surgery in childhood to repair the damage. But the surgery is not an ending. It is the beginning of a lifetime of monitoring, of regular cardiac imaging to watch for changes in heart size and pumping function that signal whether the repair is holding, whether the heart is remodeling in ways that matter. The gold standard for this surveillance has always been cardiac MRI: expensive, time-consuming, and not always easy to access. Many patients miss the scans they are supposed to have.
Researchers at Mount Sinai Kravis Children's Heart Center, supported by the National Institutes of Health, set out to change that calculus. They developed an artificial intelligence tool that reads a standard electrocardiogram—the simple, inexpensive test that takes minutes—and identifies which patients with repaired tetralogy of Fallot are at highest risk for ventricular remodeling, the kind of heart changes that typically show up on MRI. The work was published in the European Heart Journal: Digital Health.
The team trained their AI model using ECG and MRI data from patients with the condition, then tested it across five additional hospitals in North America. What they found was encouraging but not simple. The AI could estimate risk of ventricular remodeling from a routine ECG, potentially helping doctors decide which patients needed imaging urgently and which could safely wait. Performance varied from hospital to hospital, however—a finding that underscored a hard truth about AI in medicine: a tool that works well in one setting may not work the same way in another without local validation.
The implications are real. Cardiac MRI is not just inconvenient; it is a bottleneck in care. If an AI-powered ECG could help clinicians prioritize scans for the patients who need them most, while safely deferring imaging for lower-risk patients, the system could move faster and reach more people. It could also reduce unnecessary testing and the costs that come with it. For a population that requires lifelong specialized follow-up, that matters.
Son Duong, the lead author and an assistant professor of pediatrics and artificial intelligence at Mount Sinai, framed the work as an attempt to unlock hidden value from a test that already exists in every clinic. The goal is not to replace cardiac MRI—the researchers are clear on that point. MRI remains the definitive imaging tool. Rather, the AI is meant to be a triage mechanism, a way to make lifelong heart monitoring more accessible and efficient for people born with congenital heart disease.
Girish Nadkarni, a co-senior author and chief AI officer at Mount Sinai, emphasized the importance of rigorous validation across diverse clinical settings. The variation in performance across the five hospitals was not a failure; it was a lesson. Before any AI tool moves into routine clinical use, it must be tested locally, at each site where it will be deployed. The promise of AI-enabled screening is real, but so is the risk of overconfidence.
The research team plans next to run prospective clinical trials and to refine the model for younger patients, with the long-term goal of integrating the tool into standard follow-up care for congenital heart disease. If that happens, thousands of patients could benefit—not from a replacement of existing care, but from a smarter way to organize it, to get the right imaging to the right patient at the right time.
Citazioni salienti
Our goal is to make lifelong heart monitoring more accessible and efficient for people born with congenital heart disease.— Son Duong, MD, MS, lead author and Assistant Professor of Pediatrics at Mount Sinai
It is critical to rigorously validate these tools across diverse clinical settings. Our findings show both the promise of AI-enabled screening and the importance of testing performance at each site before real-world implementation.— Girish Nadkarni, MD, MPH, co-senior author and Chief AI Officer at Mount Sinai
Vuoi la storia completa?
Leggi l'originale su News-Medical.net ↗
Technology