Can an adult heart model read a child’s ECG? Fine-tuning helped—but clinical testing was small
A Chinese preprint trained on 339,563 paediatric ECGs and reached a mean AUC of 0.9695 after fine-tuning. The improvement over paediatric training from scratch was only 0.0033, and the clinician test used 100 selected cases from the same hospital.
The short version
- What happened
- The retrospective development set contained 339,563 12-lead ECGs from 238,189 children aged 0–18 at one Chinese children's hospital; an external Chinese set contained 10,999 ECGs.
- Who it affects
- Children could benefit from more consistent preliminary ECG review where paediatric expertise is limited, but the model has not been tested as an autonomous diagnostic system.
- What remains unproven
- The paper is an unreviewed preprint and the main development data come from one Chinese children's hospital.
Keep in mind
The paper is an unreviewed preprint and the main development data come from one Chinese children's hospital.
The external set is also Chinese and covers only 30 of the 45 internal labels.
The congenital-heart-disease study uses artificial 1:1 case-control balance and has only 23–28 positive cases for two subtypes.
The 100-case clinician study was selected, sequential, unblinded and conducted in the same institutional setting.
No prospective patient outcomes, prevalence-aware screening values or comparison with an ordinary second read were reported.
Source analysis · 1 linked source. Primary analysed source date: 5 November 2026.
Read the full report · 8 min The evidence, methods, implications and limitations
The direct answer: adult pretraining helped adaptation, but paediatric data did most of the work
A Chinese research team reports that a model first trained on adult electrocardiograms can be adapted to recognise paediatric ECG patterns, including uncommon labels that would otherwise be difficult to learn. Directly applying the adult model to shared labels worked poorly by comparison: its mean area under the receiver operating characteristic curve, or AUC, was 0.8256. After paediatric fine-tuning, mean AUC across 45 labels reached 0.9695 on the internal test set.
Adult pretraining was not a substitute for child data. A model with the same architecture trained from scratch on the full paediatric dataset reached 0.9661. The fine-tuned model's advantage was 0.0033 AUC, with a reported 95% confidence interval from 0.0012 to 0.0056. That is statistically detectable in a very large dataset but small in absolute terms. The practical case for pretraining is data efficiency: with only 1% of the paediatric fine-tuning set, the model reached 0.9079, and with 25% it nearly matched the full-data scratch model.[1]
What data the model actually saw
The retrospective development dataset came from the Children's Hospital, Zhejiang University School of Medicine. It contained 339,563 standard 12-lead ECG records from 238,189 children aged from birth to 18 years, collected between January 2019 and June 2023. Boys accounted for 59.9% of the children and girls 40.1%. Labels were created through the hospital's diagnostic process with two-stage review by qualified ECG diagnosticians, giving the work a clinically grounded reference rather than an automatically scraped label set.
An external dataset from Zhengzhou University included 10,999 ECGs from children aged 0–14 and supported 30 labels; the fine-tuned model's mean external AUC was 0.9385. External testing introduces a different institution, yet both sites are in China and may share acquisition practice, referral patterns and diagnostic conventions. Neonates performed worst, highlighting that age is not a minor covariate: rapid developmental changes alter rate, intervals and wave shapes in ways an adult representation may not capture evenly.[1]
The distinction that matters
Three comparisons answer three different questions
The paper's headline only makes sense when the model comparisons are kept separate.
Direct adult transfer
No paediatric adaptation
Mean AUC 0.8256
Adult then paediatric
Fine-tune on child ECGs
Mean AUC 0.9695
Paediatric from scratch
Same architecture, random start
Mean AUC 0.9661
External hospital
Test 30 labels in 10,999 ECGs
Mean AUC 0.9385
AUC is a ranking measure, not a clinical accuracy guarantee
An AUC of 0.9695 means that, across the evaluated labels, the model generally ranked a randomly chosen positive example above a negative one. It does not mean 96.95% of children were correctly diagnosed. Clinical usefulness depends on thresholds, prevalence, sensitivity, specificity and the cost of false reassurance or unnecessary escalation for each condition. Rare abnormalities can retain a high AUC while producing many false positives when deployed in a low-prevalence screening population.
The paper also tested age regression. Mean absolute error was 1.1165 years internally and 1.6602 years externally. That task helps show whether the representation encodes developmental signal, but predicting age is not itself a clinical benefit. It may capture site-specific patterns. A safer interpretation is that the representation retained age-related morphology while losing some accuracy under external shift. Prospective calibration, label-specific operating points and subgroup results would be needed before a hospital could set a reliable triage policy.[1]
The congenital-heart-disease test was promising but artificially balanced
For a downstream experiment, the team used 20,943 ECG records covering six congenital heart disease subtypes. Each subtype used an age-matched 1:1 case-control design. The model achieved a mean AUC of 0.9562. Positive cases ranged from 1,631 atrial septal defects and 1,627 ventricular septal defects to only 28 transposition-of-the-great-arteries cases and 23 coarctation cases. Results for the smallest groups rest on very few positive examples and will have wider uncertainty than the pooled headline suggests.
A 1:1 case-control set is useful for comparing representations, but it is unlike general paediatric screening, where many conditions are rare. It cannot establish real-world positive predictive value, the fraction of alerts that would truly represent disease. Nor can a surface ECG confirm structural anatomy by itself. The experiment shows that the learned signal contains information associated with diagnosed heart disease; it does not show that the model can replace echocardiography, specialist review or a prevalence-aware screening pathway.[1]
What happened when clinicians saw the AI output
The authors recruited ten full-time electrocardiographers and seven paediatric cardiologists to review 100 cases. Participants first interpreted each ECG, then saw the model's labels and Grad-CAM highlighting before answering again. Overall exact-label accuracy rose from 56.8% to 65.4%. Electrocardiographers improved by 5.2 percentage points and cardiologists by 13.5 points. The study reports that 25% of initial omissions were recovered and 18.4% of incorrect labels were corrected; 0.6% of initially correct diagnoses became wrong.
This reader study was designed to test response to assistance, not routine performance. Cases were selected for diagnostic diversity and deliberately included 19 the AI got wrong. Readers were not blinded to the second-stage intervention, always saw unaided cases first and came from the same clinical environment as the development data. When the AI was wrong, accuracy moved only from 24.5% to 25.7%, with no significant decline, but 19 cases are too few to establish resistance to automation bias across rare and consequential errors.[1]
What this could mean for children, families and clinical teams
For a child having an ECG, the potential benefit is not an autonomous diagnosis but more consistent preliminary interpretation, especially where paediatric specialists are scarce. A model that adapts with less local data could help a hospital build a paediatric tool without starting from random weights. It might flag an overlooked rhythm or morphology for review. But an alert can also create anxiety, repeat testing and referral, while a missed abnormality can delay care. Those harms depend on thresholds and local prevalence, which this paper does not evaluate prospectively.
For clinicians, the study supports designing AI as a review aid with visible uncertainty and an easy path to disagree. The larger improvement among cardiologists should not be treated as proof that experts benefit most; the case mix, initial accuracy and second-look effect can all contribute. Hospitals would need consecutive-case trials, monitoring by age and diagnosis, and a clear rule that imaging, symptoms and clinical history take priority over an ECG model's output.[1]
What would change the assessment
Confidence would rise with prospective evaluation on consecutive patients at hospitals outside China, pre-specified thresholds, prevalence-aware predictive values and comparison with ordinary second reading. A randomised or counterbalanced reader study should separate the value of AI from simply looking twice. The most important evidence would show faster, safer decisions or better patient outcomes without excessive referrals, especially in newborns and the rare diagnoses represented by only a few dozen cases here.
The assessment would weaken if calibration fails under new equipment or diagnostic conventions, if clinicians follow confidently wrong output, or if the tiny fine-tuning advantage over scratch disappears in independent replication. The work was supported by the Natural Science Foundation of China and a Zhejiang programme; the authors declare no competing interests. For now, it is a substantial retrospective representation study and a useful small simulation—not a peer-reviewed clinical validation or approval for care.[1]
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Continue the story
Read next
Continue exploring clinical evaluation: compare the finding and its evidence limits.
Can endoscopy AI travel beyond still images?
A new classifier reached about 94% accuracy in cross-validation and 96% on an aligned external image subset, then narrowly outscored five gastroenterologists on 200 still images. Both datasets came from Norway, and the work did not test live video, workflow or patient outcomes.
Read the next reportExplore the full source and evidence record
What this means for people
- Children could benefit from more consistent preliminary ECG review where paediatric expertise is limited, but the model has not been tested as an autonomous diagnostic system.
- Families could face extra referrals or false reassurance if thresholds are poorly calibrated to real disease prevalence.
- Clinicians need prospective evidence, subgroup monitoring and a workflow that makes disagreement with AI visible and safe.
Global context
Paediatric ECG interpretation is harder to scale than adult interpretation because normal patterns change rapidly with age and specialist expertise is unevenly distributed. Adult pretraining may reduce the amount of local data needed, but differences in equipment, labels, referral pathways and disease prevalence can defeat apparent portability. The useful global question is whether transparent adaptation and monitoring can improve decisions for children in different health systems without widening diagnostic inequity.
What the evidence does not yet show
- The paper is an unreviewed preprint and the main development data come from one Chinese children's hospital.
- The external set is also Chinese and covers only 30 of the 45 internal labels.
- The congenital-heart-disease study uses artificial 1:1 case-control balance and has only 23–28 positive cases for two subtypes.
- The 100-case clinician study was selected, sequential, unblinded and conducted in the same institutional setting.
- No prospective patient outcomes, prevalence-aware screening values or comparison with an ordinary second read were reported.
What to watch next
- Peer review and release of code, model details and reproducible label definitions.
- Prospective consecutive-patient validation outside China, including newborns.
- Per-label sensitivity, specificity, calibration and positive predictive value at real prevalence.
- Randomised clinician studies that separate AI assistance from a second-look effect.
- Evidence on referrals, time to diagnosis, treatment and patient outcomes rather than AUC alone.
Living evidence record
Impact record IAI-09C8OPX
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
6 November 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what arXiv published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 6 November 2026
Research question, topics and regions
Whether an adult ECG foundation model can be adapted efficiently to paediatric signals and whether its output changes specialist interpretation in a controlled reader study
Topics: Paediatric cardiology · Electrocardiograms · Foundation models · Clinical evaluation · Preprint
Regions: China · Asia · Global
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.