Artificial intelligence models designed to detect heart abnormalities from electrocardiograms work impressively well in hospitals—but put them in a sports medicine clinic, and their performance crumbles. A new study published in npj Digital Medicine reveals that three state-of-the-art ECG foundation models, despite achieving near-perfect results on clinical datasets, struggled significantly when applied to professional athletes' cardiac data, highlighting a critical "domain shift" problem that could undermine AI-powered screening programs in sports medicine.

The research tested ECG-FM, HuBERT-ECG, and ECGFounder—three foundation models pretrained on large clinical ECG datasets—against a cohort of 43 male professional soccer players. While these models achieved an impressive AUROC of 0.97 on standard clinical test data, their performance deteriorated substantially when analyzing athlete ECGs, with the best-performing model (ECGFounder) managing only an AUROC of 0.84 and a sensitivity of 0.57.

The Athlete's Heart Paradox

The performance gap stems from a phenomenon sports cardiologists know well: "athlete's heart." Years of intense training produce physiological adaptations—enlarged heart chambers, slower resting heart rates, and ECG patterns that would raise red flags in sedentary patients but represent normal, healthy responses in elite competitors. These adaptations confound AI models trained predominantly on general clinical populations.

The training dataset used for these models—PhysioNet/CinC 2021 with 87,662 ECG samples—came from patients with a mean age of 59.4 years and represented typical clinical populations from China, Germany, and the USA. The athlete cohort, by contrast, consisted of highly trained national-level soccer players with a mean age of just 20.3 years, creating a stark demographic and physiological mismatch.

ECGFounder showed the strongest performance among the three models tested, achieving its best detection rates for Sinus Bradycardia (AUROC 0.87) and Sinus Arrhythmia (AUROC 0.95)—conditions commonly seen in trained athletes. However, even these results fell short of the reliability needed for clinical deployment without human oversight.

High Stakes for Cardiac Screening

The stakes for accurate athlete cardiac screening are substantial. According to the study, ECG screening is five times more sensitive than medical history and ten times more sensitive than physical examination for detecting cardiovascular disease in athletes. Italy's mandatory cardiac screening program demonstrated the life-saving potential of systematic evaluation, reducing sudden cardiac death incidence by 90% among athletes.

Yet widespread implementation faces economic barriers. In the United States, ECG screening costs are estimated at $10.6–14.4 million per life saved—a figure that has limited adoption compared to European programs. AI-powered screening promises to reduce these costs through automation and scalability, but only if the models can accurately distinguish between benign athletic adaptations and genuine pathology.

Implications for Medical AI Development

The findings underscore a broader challenge facing medical AI: foundation models trained on general populations may not transfer effectively to specialized cohorts without domain-specific fine-tuning. For health AI companies targeting sports medicine markets, this means investing in athlete-specific labeled datasets—likely requiring hundreds to thousands of annotated ECGs from trained competitors—to achieve clinically acceptable performance.

For sports medicine physicians and team cardiologists, the message is clear: current general-purpose ECG AI models cannot yet replace expert interpretation in athlete screening. Misclassification could lead to unnecessary career terminations for healthy athletes or, more dangerously, missed detection of life-threatening cardiac conditions.

The research reveals both a technical challenge and a market opportunity—closing the domain gap between clinical and athletic populations represents a critical next step for AI developers seeking to deliver on the promise of scalable, cost-effective cardiac screening in competitive sports.

Sources

  1. ECG foundation models for cardiovascular screening in a cohort of young competitive athletes — npj Digital Medicine