AI Reads Children’s ECG Reports to Flag Hidden Heart Disease

Heart disease in children is one of the most elusive diagnostic challenges in medicine. Unlike adults, whose blocked arteries announce themselves with chest pain, children with serious cardiovascular conditions often show few or no symptoms at all, and the first clue can be easy to miss during a routine visit. Yet one of the oldest and most widely available cardiac tests, the electrocardiogram, is performed on countless children every day in hospitals around the world. A new study published in BMC Pediatrics asks a deceptively simple question: how much diagnostic information is actually hiding inside the pediatric ECG report itself, and can machine learning extract it to predict which children carry a coded cardiovascular diagnosis?

The research team, led by Pooya Eini of the Rajaie Cardiovascular Institute in Tehran alongside collaborators from Hamadan University of Medical Sciences in Iran and Michigan State University in the United States, worked with a large pediatric dataset known as ZZU-pECG. It contains 14,190 twelve-lead and nine-lead ECG recordings from 11,643 hospitalized children aged zero to fourteen years, each linked to diagnoses coded under the International Classification of Diseases, tenth revision. What makes the study distinctive is its choice of input data. In routine clinical practice, ECGs are rarely stored or exchanged as raw waveform files. Instead, they travel through hospital systems as structured reports: a list of expert diagnostic statements, patient demographics, and technical metadata. The researchers deliberately built their models on exactly this kind of report-level data, comprising 85 structured features that included the child’s age and sex, the number and identity of the expert ECG diagnostic statements attached to the recording, and per-lead signal-quality indices.

Crucially, the team excluded two variables that could have inflated the results artificially: recording duration and lead configuration. These are workflow characteristics rather than clinical signals, and including them would have allowed the models to learn shortcuts about how care is delivered rather than what the heart is doing. This kind of shortcut avoidance is a persistent worry in clinical machine learning, where algorithms have been known to achieve impressive apparent accuracy by picking up on hospital-specific artifacts rather than genuine physiology. By stripping out these workflow variables, the investigators forced their models to rely on the medical content of the report itself.

Seven different machine-learning classifiers were developed and compared, spanning the standard toolkit of modern predictive modeling: elastic-net logistic regression, random forest, histogram gradient boosting, LightGBM, support vector machine, k-nearest neighbors, and a multilayer perceptron, a small artificial neural network. Each model was trained to predict whether a child’s record carried an ICD-10-coded diagnosis from a set of 19 pediatric cardiovascular diseases. To guard against optimistic estimates, the models were tuned using three-fold cross-validation in which all recordings from the same patient stayed within the same fold, and final performance was measured on a patient-disjoint hold-out test set of 2,814 recordings. This patient-grouped design matters because multiple recordings from the same child would otherwise leak information between training and test data, making the model look better than it truly is.

The results, while promising, come with careful caveats. The best-performing model, histogram gradient boosting, achieved an area under the receiver-operating-characteristic curve of 0.821, with a 95 percent confidence interval of 0.805 to 0.837. In practical terms, that means the model could distinguish between children with and without a coded cardiovascular diagnosis reasonably well, though far from perfectly. The model was also well calibrated, meaning its predicted probabilities tracked actual outcomes closely, with a calibration slope of 1.02 and a Brier score of 0.141. Decision-curve analysis showed positive net benefit across the range of decision thresholds examined, suggesting the model could in principle add value at various levels of clinical aggressiveness.

Perhaps the most clinically meaningful numbers emerged at a high-sensitivity operating point, where the model identified 89.7 percent of recordings associated with a coded cardiovascular diagnosis, with a negative predictive value of 93.5%. In a screening context, a negative predictive value above 93 percent means that when the model says a child is unlikely to have cardiovascular disease, that reassurance is correct the vast majority of the time. For a condition that is often paucisymptomatic and easily missed at first contact, that kind of rule-out capability is exactly what a triage tool would need, flagging the small fraction of reports that deserve closer human scrutiny before they slip through the system unnoticed.

The researchers then probed where the predictive signal actually came from, and the answer is scientifically revealing. When the models were restricted to demographics and signal-quality features alone, discrimination fell to 0.748. Using only the expert ECG diagnostic statements, performance reached 0.776. Both restricted models performed significantly worse than the full model, with DeLong tests yielding P values below 0.001. Even after removing the most disease-proximal statements, the ECG statements most directly tied to a specific diagnosis, performance was largely preserved at 0.813, a statistically significant drop but a modest one. This suggests that the information content of the report is distributed across many features rather than concentrated in a few obvious diagnostic labels, and that subtler patterns in how ECG findings co-occur carry real predictive weight.

Not every subgroup fared equally well, however, and the authors are candid about these limitations. Performance dropped materially in infants, where the area under the curve fell to 0.635, and in nine-lead recordings, where it reached only 0.746. Infant ECGs are notoriously different from those of older children, with rapid heart rates and developmental patterns that make interpretation harder, and the reduced lead configuration of a nine-lead recording simply captures less cardiac information. These performance gaps, combined with the fact that the study relied entirely on internal validation within a single database, mean the models are not ready for clinical deployment. External validation on data from other hospitals, other populations, and other ECG systems would be essential before any real-world use.

There is also a deeper conceptual point that the authors emphasize: the predictors in this study are themselves expert interpretations of the ECG tracing, not the raw electrical signal. The models therefore describe the information content of the ECG report rather than of the ECG waveform itself. In one sense this is a limitation, because the models cannot exceed the diagnostic insight of the experts who wrote the original statements. In another sense it is a strength, because structured reports are what actually flow through hospital information systems today, meaning a tool built on report data could be integrated into existing workflows without requiring access to raw waveform archives that many institutions cannot easily share.

The study arrives at a moment when machine learning is being proposed for nearly every corner of medicine, and its disciplined approach offers a template for how such claims should be tested. Rather than promising a revolutionary diagnostic engine, the authors deliver a measured finding: structured pediatric ECG reports contain a reproducible, report-level signal for predicting coded cardiovascular diagnoses, strong enough to justify further research but not strong enough, yet, to change clinical practice. The work was funded by no external grants, the authors declare no competing interests, and the article is published open access under a Creative Commons license. As hospitals continue to accumulate structured ECG data on millions of children, studies like this one map out both the promise and the boundaries of what artificial intelligence can honestly extract from the paperwork of cardiology, and they remind us that the path from a promising area under a curve to a trustworthy bedside tool runs through external validation, subgroup analysis, and a healthy respect for what the data can and cannot say.

Subject of Research: Machine-learning prediction of pediatric cardiovascular diagnoses from structured electrocardiogram report data

Article Title: Machine-learning prediction of cardiovascular diagnoses from pediatric ECG reports in 11 643 children

Article References: Eini, P., serpoush, H., Rezayee, M., & Tremblay, J. (2026). Machine-learning prediction of cardiovascular diagnoses from pediatric ECG reports in 11 643 children. BMC Pediatrics. https://doi.org/10.1186/s12887-026-07732-3

Image Credits: AI Generated

DOI: 10.1186/s12887-026-07732-3

Keywords: machine learning, pediatric ECG, cardiovascular disease, clinical prediction model, electrocardiography, predictive medicine, children's health, gradient boosting, ICD-10 coding, diagnostic screening, Machine-learning, prediction