AI Analysis of Routine ECGs May Help Prioritize Cardiac MRI in Congenital Heart Patients

Tetralogy of Fallot is the most common complex congenital heart defect, affecting roughly 1 in 2,500 live births. Surgery in infancy or early childhood can correct the anatomical problem, but the repair is not a cure. The right ventricle continues to face abnormal hemodynamic stress throughout life, and over decades that stress produces structural changes - ventricular remodeling - that can eventually impair heart function and require further intervention.

Cardiac MRI is the standard tool for tracking these changes. It measures right ventricular size and function with precision that echocardiography cannot match. But MRI is expensive, time-consuming, not universally accessible, and carries particular logistical challenges for patients who live far from specialized centers or who require sedation. Many patients with repaired tetralogy of Fallot miss their recommended imaging. The result is that some develop significant ventricular dysfunction that goes undetected until symptoms become severe.

Training an AI on ECG Patterns

A multicenter team led by researchers at the Mount Sinai Kravis Children's Heart Center trained an artificial intelligence model to identify ECG patterns linked to ventricular remodeling in repaired tetralogy of Fallot patients. The ECG is inexpensive, takes minutes, and is available in nearly any clinical setting. If AI could extract information from ECG signals that correlates with the changes cardiac MRI detects, it could serve as a pre-screening tool - identifying patients who need MRI most urgently while safely deferring imaging in lower-risk individuals.

The training dataset combined ECG recordings and contemporaneous MRI measurements from patients with repaired tetralogy of Fallot. The AI model learned which ECG features - patterns in the electrical signals produced by the heart's various chambers - associate with the kinds of ventricular size and function changes that indicate worsening health. The study was supported by the National Institutes of Health and published in the European Heart Journal: Digital Health.

Multicenter Validation

The team validated the model across five additional hospitals in North America, testing whether its performance held in clinical settings with different equipment, patient populations, and ECG recording protocols than those used in training. The model demonstrated the ability to identify patients at risk for ventricular remodeling, with potential to help clinicians prioritize MRI for higher-risk patients and safely delay scans for lower-risk ones.

An important caveat emerged in the validation process: performance varied by hospital site. This finding does not undermine the concept, but it does underline a critical principle for clinical AI deployment. A model trained on data from one institution may not generalize automatically to another with different patient demographics, equipment characteristics, or clinical practices.

"As AI becomes more integrated into health care, it is critical to rigorously validate these tools across diverse clinical settings," said Girish Nadkarni, co-senior author and Chief AI Officer at Icahn School of Medicine at Mount Sinai. "Our findings show both the promise of AI-enabled screening and the importance of testing performance at each site before real-world implementation."

What the Tool Is and Is Not

"This research shows how artificial intelligence can extract new value from a routine ECG," said Son Duong, lead author and Assistant Professor of Pediatrics at Mount Sinai. "Our goal is to make lifelong heart monitoring more accessible and efficient for people born with congenital heart disease."

The researchers are explicit that the tool is not designed to replace cardiac MRI. Ventricular remodeling requires MRI for definitive quantification - the ECG-based model provides a risk estimate, not a measurement. The intended application is triage: identifying which patients in a large population of repaired tetralogy of Fallot adults most urgently need their next MRI scan, and which can safely wait.

That triage function has real practical value. The population of adults with repaired congenital heart disease is growing as surgical outcomes have improved, and specialized adult congenital heart programs are stretched. A validated risk-stratification tool that reduces unnecessary imaging while catching high-risk cases sooner could meaningfully improve both access and efficiency.

Next Steps and Limitations

The research team plans prospective clinical studies to evaluate the tool in real-world use rather than the retrospective analysis that characterized this initial validation. They also plan to refine the model for younger patients, who may have different ECG characteristics than the adults who comprised most of the study cohort. The long-term goal is integration into routine clinical workflow.

The site-to-site performance variation observed in validation suggests that local calibration may be necessary before deployment at any specific institution - an implementation requirement that adds complexity to rollout. The model was also validated specifically for tetralogy of Fallot; whether the approach generalizes to other congenital heart defects remains untested.