This page explains the research in order. Start with the main idea and the plan (Figure 1), then see where the research stands . Parts 1–3 show what is done: (1) transferability testing , (2) normal ECG ranges by age and (3) the age-guided ECG simulator . Part 4 lists the next steps. The sources behind the 3D anatomy explorer are in the appendix at the end.
Main idea We will use knowledge of how children’s hearts change as they grow to generate synthetic ECGs for training. We will test whether these examples help models detect arrhythmias in real children better than examples generated using age information alone.
Research question Compared with using age alone, can adding knowledge of heart development and physiological simulation help synthetic ECG training data improve arrhythmia detection across children’s ages and rhythms, reduce the amount of labeled pediatric data needed, and better identify ECGs that should receive priority clinical review?
View full figure ↗ Figure 1. Proposed Heartverse framework for developmental physiology-guided ECG augmentation and pediatric arrhythmia detection.
(A) Study design. Pediatric ECGs are standardized, and patients are separated into training, validation, and test groups before recordings are divided into segments. Adult-to-child transfer is examined as a supporting study. Generator A uses age information, while generator B also uses knowledge of how the heart changes during development and physiological simulation. Both generators use the same real training patients, model architecture, and training budget. The same classifier architecture is trained separately using real ECGs only, real ECGs plus samples from A, or real ECGs plus samples from B. Standard data augmentation is included for comparison.
(B) ECG generation. Both generators use exact age, rhythm labels, and the same patient information. Generator B additionally uses normal training ECGs and published age references to guide simulation while preserving the intended abnormal rhythm. Synthetic ECGs are added either according to a preset age–rhythm mix or to groups where the model makes more errors. These difficult groups are identified using validation within the training data. Within each strategy, A and B generate the same number of ECGs with the same age–rhythm mix.
(C) Evaluation. Performance will be measured separately by age group and rhythm using real patients kept out of training and model adjustment. Every step that uses pediatric labels will follow the same label budget. We will also examine how the models could prioritize ECGs for review, measuring missed cases, false alarms, and review workload. This analysis will use the actual proportion of positive cases in the test data and decision thresholds chosen using validation data.
Ages 1-16 years are the proposed scope; final age coverage and group boundaries depend on available patients and reliable rhythm labels. Exact age guides generation, while age groups summarize evaluation; axis ticks indicate age values rather than fixed group boundaries. The figure depicts a proposed study, and its signals, distributions and model blocks are schematic rather than empirical results. Demonstrating clinical decision benefit requires a clinician reader study or prospective evaluation.
Where the research stands
The work moves in order. Each part builds on the one before it. Click a part or a figure number to jump to it.
Part What it answers Figures Status
1. Transferability testing Does an ECG model trained on adults still work for young adults, children and infants? 2 , 3 , 4 Done 2. Normal ECG ranges by age What does a normal ECG look like at each age from 1 to 14? 5 , 6 Done 3. Age-guided ECG simulator Can we make synthetic ECGs that look their age and keep the intended rhythm? 7 , 8 , 9 , 10 Done 4. Generators A and B Does adding developmental profiles and simulation to diffusion make better synthetic ECGs? 1 (plan)In progress 4. Classification study Do synthetic ECGs help detect arrhythmias in held-out real children? 1 (plan)Planned
PART 1 · DONE
Transferability Testing Question: does an ECG model trained on adults still work for young adults, children and infants?
Study design and results for testing how ECG conduction classification transfers across age groups.
Key finding: a model trained only on adults aged 40–89 worked on independent young adults without retraining (AUROC 0.901 ). In children, giving the adult model extra pediatric training worked best (AUROC 0.941 , versus 0.870 from scratch). Infant results are still uncertain.
Preparation
Selected one task: detecting conduction abnormalities using two classes—conduction-positive and conduction-negative .
Matched labels across datasets: we used a common label definition. The mapping still needs clinician review.Defined age groups: older adults 40–89 , young adults 20–29 , children 1–14 , and infants under 1 year .Standardized the ECGs: we used 12 leads, 10-second recordings, and a 100 Hz sampling rate , with consistent voltage units.Separated patients: the pediatric analysis used one ECG per patient and five patient-level folds . Each patient was tested by models that had not used that patient for training or adjustment.Evaluated real ECGs: we reported results separately for each age group. Infant findings remain uncertain because there were only 20 positive cases .
View full figure ↗ Figure 2. Study design for evaluating age transfer in ECG conduction classification. An older-adult-trained model was evaluated without adaptation on an independent young-adult cohort. Pediatric experiments compared older-adult and young-adult fine-tuning with training from scratch using identical pediatric training patients. Five patient-level folds kept test patients separate from training, model selection, calibration, and threshold selection. Young-adult evaluation assessed direct transfer, while exploratory pediatric experiments assessed the benefit of adult pretraining.
Results: ECG classification across age groups
We measured performance using AUROC , which shows how well a model separates conduction-positive from conduction-negative ECGs. Higher values are better: 0.5 means chance-level performance, and 1.0 means perfect separation .
A. Older adults → young adults
The older-adult model was tested on an independent young-adult dataset without additional training or adjustment .
It achieved:
AUROC: 0.901 , with a 95% confidence interval of 0.857–0.939 .Sensitivity: 70.6% —the percentage of positive cases correctly identified.Specificity: 91.5% —the percentage of negative cases correctly identified.
The same model achieved AUROC 0.893 in the external older-adult group.
These results support direct transfer: the model used what it had learned from older adults to make useful predictions in young adults. The similar scores do not prove that performance was exactly equal in both age groups.
B. Transfer to children
We tested the models on 9,842 children aged 1–14 , including 710 conduction-positive cases .
Training approach AUROC
Older-adult model given additional pediatric training 0.941 Young-adult model given additional pediatric training 0.874 New model trained using pediatric data alone 0.870
Giving an existing model additional training is called fine-tuning . Training a new model without previous adult training is called training from scratch .
Older-adult fine-tuning improved AUROC by 0.071 compared with training from scratch. The estimated 95% uncertainty range for this improvement was 0.058–0.086 , entirely above zero.
At the selected decision thresholds:
The adapted older-adult model detected 590 of 710 positive cases .
The model trained from scratch detected 488 of 710 .
They produced similar numbers of false positives: 887 versus 894 . These are negative cases incorrectly flagged as positive.
Both approaches used the same pediatric training patients. This supports the benefit of earlier older-adult training under our study conditions.
Starting with the young-adult model showed no clear extra benefit. Its AUROC improvement was 0.004 , with an uncertainty range of −0.004 to 0.013 , which includes zero.
C. Transfer to infants
We tested the models on 249 infants , including 20 conduction-positive cases .
Training approach AUROC Positive cases detected False positives
Older-adult model given additional pediatric training 0.834 11/20 14 Young-adult model given additional pediatric training 0.723 9/20 22 New model trained using pediatric data alone 0.736 7/20 14
The adapted older-adult model detected four more positive infants than the model trained from scratch, with the same number of false positives.
However, it still missed 9 of 20 positive cases . The uncertainty range for its AUROC improvement was −0.003 to 0.215 . Because this includes zero, we cannot confidently confirm an improvement.
These results suggest that older-adult training may help with infant ECGs, but more evidence is needed. Reliable infant screening has not been established.
Our findings support transferability for this ECG conduction-detection task . A model trained only on adults aged 40–89 performed well on 2,138 independently collected young adults aged 20–29 , achieving AUROC 0.901 without additional training or adjustment. This shows that what it learned from older adults remained useful in young adults. In 9,842 children , giving the older-adult model pediatric training achieved AUROC 0.941 , compared with 0.870 for a new model trained on the same pediatric patients. The estimated improvement was 0.071 , with a descriptive 95% interval of 0.058–0.086 , supporting a benefit from older-adult pretraining under our study conditions. Starting with the young-adult model showed no clear extra benefit. In infants, the adapted older-adult model detected 11 of 20 positive cases , compared with 7 from scratch , with 14 false positives each . However, it still missed nine cases, and the estimated improvement remained uncertain. Therefore, our strongest evidence is for direct transfer to young adults and improved childhood performance after adaptation . The pediatric findings come from exploratory five-fold evaluation on previously examined data and still need confirmation in a new independent dataset.
PART 2 · DONE · BUILT ON REAL CHILDREN'S ECGS
Normal ECG ranges by age Question: what does a normal ECG look like at each age from 1 to 14?
A child's ECG changes a lot as they grow. Before we can make synthetic ECGs that look their age, we need to know what a normal ECG looks like at each age. So we built age-specific profiles from real children's ECGs.
Key finding: the profiles reproduce how a normal ECG changes from age 1 to 14 (heart rate falls, QT lengthens, right-ventricular dominance fades), and their centiles are well calibrated (1.9–2.6% of children beyond p2 or p98).
How we built the profiles
Used real ECGs: 12-lead ECGs from the ZZU pECG dataset, children aged 1–14 .Kept only clearly normal ECGs: the ECG is labeled normal, has 12 leads and at least 10 seconds, has good signal quality (bSQI ≥ 0.9 ), and the child has no heart-related or rate-affecting diagnosis code. This strict-normal cohort has 799 children , one ECG each.Locked away a test set: 2,370 patients (20% of ZZU) are held out for final testing. This step never used them.Measured each ECG: heart rate, PR, QRS, QT, axes and wave amplitudes, taken from each child's median beat.Fitted smooth age curves: for each measurement we estimated the centiles (p2, p25, p50, p75, p98) by age and sex. The measurements are also kept consistent with each other, so a sampled child is realistic as a whole.Infants under 1 year: we use published reference values only (Rijnbeek 2001; PHN 2018), because there were too few normal infant ECGs.
Findings
The profiles reproduce the known changes of childhood , and the measurements were checked in several ways.
A. What changes with age
Measurement Age 1 Age 14 What it means
Heart rate (p50) 132 bpm 76 bpm The heart slows down. QT interval (p50) 282 ms 348 ms Each beat takes longer. R wave in V1 (p50) 0.98 mV 0.39 mV The right ventricle becomes less dominant (right-ventricular dominance fades). R wave in V6 (p50) 0.54 mV 1.23 mV The left ventricle takes over. T wave upright in V1 0% (1–2 y) 33% (10–14 y) The T wave in V1 slowly turns upright. QT–RR exponent 0.52 (1–2 y) 0.37 (10–14 y) Neither Bazett (0.5) nor Fridericia (0.33) fits every age.
B. How we checked the profiles
Check Result
Are the centiles calibrated? Yes. Out-of-fold, 1.9–2.6% of children fall below p2 and above p98 for every measurement (about 2% expected). Do the measurements match the ZZU diagnoses? 11 of 12 checks pass. For example, PR separates first-degree AV block with AUROC 0.99 , and QTc separates prolonged QTc with AUROC 0.92 .Do they agree with published values? Heart rate, P duration and QTc agree within about ±8% . PR is 7–9% longer and QRS about 17% shorter, because of how the start and end of each wave are measured. Do chest-lead amplitudes transfer? Not fully. Chest-lead amplitudes differ between hospitals and devices (R in V6 is 0.61–0.77× published), so amplitude is treated as a recording setting, not physiology.
Limits: only 34 children aged 1–2 , so the youngest centiles are wide. The data come from one hospital and are cross-sectional: they compare different children at different ages, not one child growing up. These are research profiles, not clinical reference ranges.
Sources: ZZU pECG (Sci Data 2025 ); Rijnbeek et al. 2001 (Eur Heart J ); Saarel et al. 2018, Pediatric Heart Network (Circ Arrhythm Electrophysiol ).
PART 3 · DONE · GUIDED BY THE AGE PROFILES
Age-guided ECG simulator Question: can we make synthetic ECGs that look their age and keep the intended rhythm?
The simulator makes synthetic 12-lead ECGs (10 seconds, 500 Hz) for a child of a chosen age (1–14) and sex. Because it draws from the age profiles above, each ECG looks its age. It can also add a chosen rhythm, so every synthetic ECG comes with an exact rhythm label .
Key finding: synthetic ECGs mostly look their age (84–95% inside the age p2–p98 range) and keep the intended rhythm (85–99.5% ), but a model can still tell them apart from real ECGs (AUROC 0.88 ).
How the simulator works
Step What it does
1. Timing Draws heart rate, PR, P duration, QRS, QT and beat-to-beat variability (RMSSD) together from the age profiles. 2. Heart vector (VCG) Builds the heart's electrical direction in 3D (a vectorcardiogram ) from 8 smooth waves: 2 for P, 4 for QRS and 2 for T. 3. Shape by age Wave shapes and sizes come from an age model fitted to 783 strict-normal ZZU children . 4. 12 leads Projects the 3D heart vector onto the 12 standard leads, using each child's own lead fields (inverse Dower matrix). 5. Rhythm Normal sinus rhythm, sinus tachycardia (above the age p98), sinus bradycardia (below the age p2), or 1–3 early beats: PAC or PVC . 6. Recording noise Adds baseline wander, broadband noise and 50 Hz mains interference.
Findings
Synthetic ECGs mostly look their age and keep the intended rhythm . They are not yet indistinguishable from real ECGs.
Check Result
Can it rebuild real children? Yes, closely. For 200 real children rebuilt from their own fitted settings, QRS is off by +2 ms , QT by +8 ms , and the heart axes by 5° or less . Do normal ECGs look their age? Mostly. 84–95% of synthetic normal ECGs fall inside the age p2–p98 range, compared with 96% for real children. The T axis is weakest at 77%: its spread is too wide. Is the rhythm kept? Yes. Tachycardia 90.5% , bradycardia 99.5% , PAC 93.1% and PVC 85.3% . PVCs are 2.1× wider than the child's own normal beats. Are the normal beats in PAC/PVC ECGs still age-appropriate? Yes. 83–92% stay inside the age p2–p98 range for PR, QRS, QT and QRS axis. Can a model tell real from synthetic? Yes, still: AUROC 0.88 (0.5 would mean they cannot be told apart). Before our fixes it was 0.99. T axis and P duration give them away most. Can age be read back from a synthetic ECG? Nearly as well as from a real one: average error 2.1 years for synthetic children versus 1.8 years for real children.
How it is used next: the simulator is not the final generator. It gives generator B its developmental knowledge and exact rhythm labels, and its checks (Figure 10) will be reused to judge both generator A and generator B.
Limits: ages 1–14 only, with no infant model yet. PAC and PVC shapes follow hand-made rules; they have not yet been compared with real early beats or reviewed by a clinician. Beat-to-beat variability is simple, with no breathing-related changes, and noise levels are assumed rather than measured.
Source: inverse Dower coefficients from Vondrák et al. 2022 (PMC9106114 ).
PART 4 · NEXT
Next steps These steps complete the main study planned in Figure 1 . Steps marked Done are covered in Parts 2 and 3.
Done · see Figures 5–6 We will start by building age-specific profiles using pediatric references and normal ECGs from our training data to describe how heart rate, intervals, and waveforms vary as children grow.Done · see Figures 7–10 Those profiles will guide our simulator in generating age-appropriate ECGs while preserving the intended arrhythmia.In progress · generator A is training Our main comparison will use two generators:
A: Age-conditioned diffusion.B: The same model with additional developmental profiles and simulation guidance.Before adding synthetic ECGs to training, we will assess age plausibility, rhythm preservation, waveform quality, diversity, and copying from the training ECGs.For augmentation, we plan to compare a planned sample mix with extra examples for difficult age–rhythm groups. Synthetic counts will be matched, with training-internal validation identifying where our models struggle.The classification study will compare real data alone, conventional augmentation, and synthetic augmentation on held-out real pediatric ECGs.To measure how much labeled data we need, we will repeat the experiments with 10%, 25%, 50%, and 100% of pediatric labels. Our generators and classifiers will share the same label budget.We also want to assess review prioritization by comparing how many arrhythmias are detected at a fixed review capacity and how many false alarms occur at matched sensitivity.Our final analysis will examine which components help, separating the effects of developmental profiles, simulation, and targeted augmentation and reporting results for each age and rhythm group.
APPENDIX · ANATOMY EXPLORER
References & assumptions Sources for the size measurements, anatomy, and motion shown in the explorer.
3D model: Human_heart_3d_model by marxtrax · CC BY 4.0 . Adapted materials, lighting, color contrast and visual scale.
What changes when you move the slider?
How does the 4D cycle work? The cycle scrubber plays a gentle illustrative deformation of this existing mesh. Its timing follows the measured size of the Time2Track ground-truth mesh in each of the 49 released frames: projected area, smoothed over five frames, with frames assumed evenly spaced over one heartbeat because the source has no frame durations. The heart starts contracting after frame 3 and stays near its smallest size from frame 16 to 27. Projected area is a size proxy, not a measured volume, and the amount of deformation is illustrative. It does not change the selected published LVEDV reference. Rotation and zoom only change the view. The outline stays at the selected sex’s BSA 1.0 m² reference for children, or the 143 mL pooled mean for adult men. No patient-specific contraction, pressure or infant-transfer model is present.
How does the sample ECG work? The lead II strip and the 3D loop are a sample ECG built from published median heart rate, PR, P and QRS duration, Bazett-corrected QT and frontal axes for the selected sex and ECG age group (Rijnbeek 2001 for children aged 1–16; Rijnbeek 2014 for adult men). At 1× the 4D cycle beats at that median heart rate; slower and faster settings are playback speeds, so measured intervals stay the same. The QRS starts just before the illustrative contraction and the T wave ends as relaxation begins. The 3D loop is a vectorcardiogram: the heart’s electrical direction traced through one beat. The strip is that loop projected onto the lead II axis, shown as the dashed line. Wave shapes, front–back loop components, T-wave size, adult amplitudes and the electromechanical delay are illustrative. The ECG is not computed from the 3D mesh, and heart size does not change it. It is not a patient ECG and not diagnostic.
What does the motion reference show? The panel replays a released shift-invariant baseline comparison from the Time2Track repository, shown as released: the left mesh is that baseline model’s output and the right mesh is the source’s ground truth. Heartverse makes no predictions of its own. It preserves all 49 frames, which the repository describes as one generated cardiac cycle. With Sync with heart on, frame N is shown at the same moment of the 4D cycle (N/49 of the beat), and the panel’s play button and frame slider drive the 3D heart and sample ECG too; the beat rate comes from the selected ECG age group, not from the source. With sync off, it plays on its own at an adjustable display rate that is not measured heart rate. The example is independent of the size controls, and the site does not run the ECG model. Use the expand button to inspect individual frames.
Where the evidence stops A cavity volume does not define a unique 3D heart. The source mesh is an artistic anatomical cutaway with its own proportions. Its cavity volume is not measured here. A change in size is not an age prediction. Statistical summaries from different cohorts are not interchangeable. The pediatric reference covers ages 4–18. The adult reference is a separate pooled population, not an extension of the childhood curve. Neither supports infant extrapolation or adult-to-child algorithm transfer. This explorer does not accept individual patient data. Its separate motion panel displays a released research example.
References checked October 14, 2026. Full data and assumptions are included in the downloadable reference file.
↓ Download reference & current setting