BeatGraph: Self-Supervised Heartbeat Graphs for Infant ECG Representations from the Home Environment

1 INTRODUCTION

• Existing ECG models mainly use adult clinical recordings, while infant ECG differs in heart rate, morphology, intervals, and wearable lead configuration.Source: §1 · p. 2
• F1-score of 0.589 to 0.923 across four infant tasks, comparable performance on ZZU-pECG, and a match to ST-MEM on PTB-XL under linear evaluation are reported for the infant-pretrained encoder.The introduction notes that the encoder trails after full fine-tuning on PTB-XL.Source: Figure 1 · p. 2, §1 · p. 2
• Fixed-length patches can split heartbeats and contain varying beat counts, requiring a model to recover cardiac structure before relating beats.Source: §1 · p. 2
• BeatGraph represents each 30-second ECG window as a graph of beats, encoding waveform and intervals before temporal ordering, graph attention, and window-level pooling.It is pretrained by masking beats and predicting their embeddings with a self-distillation regularizer.Source: §1 · pp. 2–3, Figure 2 · p. 3
• With adult ECG pretraining, the same architecture leads all compared methods under linear evaluation, reaching 0.825 against 0.804 for ST-MEM in the single-channel setting.Source: §1 · p. 3
• The released corpus contains 3,408 hours of ECG from 143 infants aged 3 to 11 months, including unlabeled pretraining data and labeled state and affect benchmarks.The authors describe it as the first public infant ECG corpus collected in naturalistic settings with behavioral-state and affect labels.Source: §1 · p. 3

ECG foundation models

• ECG foundation models target adult clinical multi-lead data and tokenize waveforms into patches or leads [Na et al., 2024; Nguyen et al., 2025; Liu et al., 2024].Examples include ST-MEM’s spatio-temporal masked reconstruction and ECG-FM’s contrastive and generative objectives [Na et al., 2024].Source: §2 · pp. 3–4

Self-supervised and graph learning for physiological signals

• Self-supervised physiological-signal methods use contrastive, self-distillation, or masked modeling, while ECG methods commonly operate on samples, segments, or patches.Source: §2 · p. 4
• Prior graph-based ECG methods build graphs from signal samples, engineered-feature correlations, or topological transforms, typically for supervised arrhythmia classification [Oliveira et al., 2024; Han, 2024; Zeinalipour & Gori, 2023].Source: §2 · p. 4

Pediatric and infant data

• Public pediatric ECG datasets are clinical and lack developmentally relevant state labels, whereas BeatGraph’s wearable data support evaluation across sleep, autonomic, behavioral, and affective states.BeatGraph makes individual beats the tokens, leaves the edge set unrestricted, and uses masked beat-embedding prediction with BYOL-style self-distillation on unlabeled single-channel infant recordings.Source: §2 · p. 4

3 METHOD

• BeatGraph combines a shared beat encoder, a temporal Transformer, and graph attention to produce one window-level embedding.Attention-weighted pooling produces the final embedding.Source: §3 · p. 4, §3.2 · pp. 4–5

3.1 BEAT-GRAPH CONSTRUCTION

• Each beat is centered on an R-peak, carries waveform and inter-beat timing information, and is represented as a node in a complete graph of valid beats.Windows are padded to at most 80 beats, with masks excluding padded positions from attention, pooling, and losses.Source: §3.1 · p. 4

3.2 ENCODER ARCHITECTURE

• A shared convolutional encoder combines pooled waveform features with timing features, then a positional Transformer orders beats before three residual graph-attention layers refine them.The graph-attention heads are concatenated, projected, and added residually; attention-weighted pooling forms the window embedding.Source: §3.2 · pp. 4–5

3.3 SELF-SUPERVISED PRETRAINING

• Pretraining predicts embeddings for 30% of valid beats masked before the Transformer, using an unmasked, gradient-stopped target pass and a BYOL self-distillation regularizer.The BYOL regularizer uses a symmetric cosine loss between online and exponential-moving-average target networks.Source: §3.3 · p. 5

3.4 DOWNSTREAM FINE-TUNING

• For each task, the pretrained encoder is fine-tuned with a residual-MLP head on 30-second windows; external-data evaluation also includes frozen-encoder linear evaluation.For multi-lead external datasets, the single-channel encoder is applied to each lead and the embeddings are concatenated before classification.Source: §3.4 · p. 5

3.5 IMPLEMENTATION DETAILS

• The encoder has 1.37 million parameters, and pretraining runs for 100 epochs with AdamW, learning-rate warm-up and cosine decay, and an EMA target decay of 0.99.The BYOL target network and projection heads are discarded after pretraining.Source: §3.5 · p. 6

4.1 SENSING PLATFORM AND COHORTS

• The corpus includes 3,361 hours of unlabeled recordings from 139 infants and 39.3 hours of labeled classroom recordings from four infants.The affect task uses laboratory playroom recordings from 122 infants in the first cohort.Source: §4.1 · p. 6

4.2 PRETRAINING DATA

• Pretraining uses approximately 3,361 hours of unlabeled home ECG; labeled classroom recordings come from a disjoint cohort, and affect sessions are separate laboratory recordings.No labeled recording enters the pretraining corpus.Source: §4.2 · p. 6

4.3 DOWNSTREAM TASKS

• The first three tasks use leave-one-infant-out splits; affect evaluation tests generalization to unseen labels, since held-out infants contribute unlabeled pretraining data.Source: §4.3 · p. 6
• The four tasks classify sleep versus wake, four infant states, infant- versus caregiver-initiated movement, and positive versus negative affect using play and still-face episodes as proxies.The four-state labels are crying/fussing, quiet alert, active alert, and sleeping; movement labels distinguish infant-generated activity from caregiver-influenced movement.Source: §4.3 · pp. 6–7

4.4 RELEASED BENCHMARK

• The released resource includes raw and processed ECG, task definitions, subject-level splits, and code; it is presented as the first public infant ECG resource with naturalistic continuous recordings and multiple behavioral annotations.It covers sleep/wake, infant states, activity source, and affect in children aged 3 to 11 months; affect labels come from lab visits, while the other annotations use home and classroom recordings.Source: §4.4 · p. 7

4.5 EXTERNAL DATASETS FOR TRANSFER EVALUATION

• Transfer evaluation uses ZZU-pECG with 14,190 records from 11,643 children aged 0 to 14, and PTB-XL with 21,799 adult records from 18,869 patients.MIMIC-IV-ECG provides adult pretraining data, with PTB-XL held out for evaluation.Source: §4.5 · p. 7

• Classroom evaluation used leave-one-infant-out folds, whereas affect used subject-level splits across 122 dyads despite pretraining exposure to those infants.Thus, affect results measure generalization to unseen labels rather than unseen subjects.Source: §5 · pp. 7–8
• BeatGraph achieves the strongest accuracy, macro-F1, and Cohen’s κ across all four infant tasks against methods pretrained on the same infant corpus.Source: Table 1 · p. 7, §5 · p. 8
• BeatGraph-Infant matches ST-MEM under PTB-XL linear evaluation despite pretraining only on single-channel ECG from infants aged 3–11 months.Both BeatGraph variants are stronger frozen than fine-tuned relative to ST-MEM, which the authors interpret as gains from representation quality rather than adaptation capacity.Source: Table 2 · p. 8, §5 · p. 8
• 0.129 AUROC is the twelve-to-one-lead loss for ST-MEM, versus 0.106 for BeatGraph-Adult; BeatGraph-Adult leads by 0.021 in the single-lead setting.With twelve leads, BeatGraph-Adult reaches 0.931 ± 0.004 against ST-MEM’s 0.933 ± 0.003.Source: Table 3 · p. 8, §5 · p. 8
• 0.892 AUROC places infant-pretrained BeatGraph within the 0.883–0.893 range of four models pretrained on large multi-lead clinical corpora on ZZU-pECG.ZZU-pECG is hospital-recorded diagnostic ECG, so the comparison spans recording settings as well as age.Source: Table 4 · p. 8, §5 · p. 8
• Adult-pretrained ECG checkpoints perform weakly when directly applied to infant tasks after Lead I placement and zero-filling the other leads.The authors conclude these checkpoints cannot be used directly on infant wearable ECG, whether because of pretraining domain or lead adaptation.Source: §5 · p. 8, Table 5 · p. 9
• Graph attention beats deeper Transformer-only variants, heartbeat alignment helps consistently, and RR features add uneven benefit across tasks.The fully connected graph performs best on three of four tasks, while 32-kNN is slightly better for activity-source recognition. A 40-beat cap truncates almost every 30-second window; 120 mostly adds padding, and detector substitution has a comparable effect across tasks.Source: Table 6 · p. 9, Table 7 · p. 9, Table 8 · p. 9

6 CONCLUSION

• BeatGraph outperforms waveform-based baselines on four infant tasks, remains competitive on ZZU-pECG, and matches adult foundation models on PTB-XL under linear evaluation.The release includes 3,361 hours of unlabeled infant ECG, 47.5 annotated hours, subject-level splits, and task definitions.Source: §6 · p. 9

A.1 COMPOSITION BY RECORDING SETTING AND LABEL SOURCE

• Home recordings provide 3,361 hours of unlabeled pretraining ECG from 139 infants, while classroom annotations cover 39.3 hours from four infants and laboratory affect labels cover 8.2 hours from 122 dyads.The laboratory labels use play and still-face episodes as proxies for positive and negative affect.Source: §A.1 · p. 14
• The resource combines home pretraining data with behavioral annotations from daycare and laboratory subsets; it is not a home-labeled corpus.“In the wild” applies only to home and classroom recordings, not still-face sessions.Source: §A.1 · p. 14

A.2 ETHICAL REVIEW AND CONSENT

• Every recording followed institutional review board approval and written guardian consent, including consent for public release of deidentified physiological data.Families could withdraw and request deletion.Source: §A.2 · p. 14

A.3 DEIDENTIFICATION AND RE-IDENTIFICATION RISK

• Long continuous recordings and released beat timing increase matching material, while changing infant cardiac morphology and variable electrode placement limit template stability.Residual linkage risk remains for someone who already has ECG from a specific infant in the cohort.Source: §A.3 · pp. 14–15

A.4 WHAT IS AND IS NOT RELEASED

• The release includes raw 1000 Hz ECG, processed 30-second windows, R-peak locations, inter-beat intervals, task labels, subject-level splits, and demographic variables; audio, IMU, and video are withheld.Raw and processed forms allow users to apply their own beat detection.Source: §A.4 · p. 15

A.5 DEMOGRAPHICS AND REPRESENTATION LIMITATIONS

• The cohort covers ages 3–11 months, three benchmark tasks rely on four classroom infants, and all recordings use one wearable device.Affect labels come from 122 laboratory dyads but use episode identity as a proxy rather than direct affect annotation.Source: §A.5 · p. 15

A.6 INTENDED USE, MISUSE POTENTIAL, AND LIMITATIONS

• BeatGraph is a research artifact, not a diagnostic device, and has not been validated for clinical diagnosis, screening, monitoring, or care decisions.The external benchmarks measure representation transfer, not diagnostic fitness.Source: §A.6 · p. 15
• Caregiver-initiated movement carries no validity as a measure of caregiving quality, and Still-Face affect proxies do not support claims about ordinary-setting emotional states.The paper cautions against presenting model outputs as assessments of individual infants or caregivers.Source: §A.6 · p. 16
• Three task labels come from four infants, affect labels are episode proxies, affect-task held-out infants contribute unlabeled home data to pretraining, and beat detection lacks manual validation.The affect result measures generalization to unseen labels rather than unseen subjects.Source: §A.6 · p. 16

B.1 WHICH BASELINES ARE COMPARABLE, AND WHERE

• Infant-task baselines must support single-channel data and scratch pretraining; lead-geometry objectives are excluded or replaced by temporal counterparts where applicable.For external twelve-lead benchmarks, the authors use the published baseline sets and add methods they reran.Source: §B.1 · p. 16

B.2 BASELINES PRETRAINED ON THE INFANT CORPUS

• Infant-task methods share the 3,361-hour unlabeled split, 30-second windows, fine-tuning schedule, subject partitions, and matched-capacity heads, with no labeled recordings in pretraining.This setup is intended to make encoder differences the source of performance differences.Source: §B.2 · p. 16
• The baseline set includes ECG-FM with random lead masking removed for single-channel use, HuBERT-ECG masked-unit prediction, and SimCLR contrastive learning on augmented views.ECG-FM retains its first two loss terms; HuBERT-ECG uses two clustering iterations.Source: §B.2 · p. 16
• BYOL serves as a baseline for separating beat-graph tokenization from the self-distillation regularizer also used in BeatGraph.Its online network trains against an exponential-moving-average target without negative pairs.Source: §B.2 · p. 17
• MTAE is the closest patch-based counterpart to BeatGraph and provides the comparison intended to isolate tokenization.It masks temporal patches and reconstructs them with a lightweight decoder.Source: §B.2 · p. 17

B.3 ADDITIONAL BASELINES FOR EXTERNAL TRANSFER

• External transfer compares BeatGraph against published baseline sets on PTB-XL and ZZU-pECG, adding two methods the authors reran themselves.The listed comparisons include MoCo v3, CMSC, MTAE, MLAE, ECG-FM, TolerantECG, ST-MEM, MERL, and HuBERT-ECG.Source: §B.3 · p. 17
• Table 11 distinguishes scratch-pretrained, published, and reproduced results; ECG-FM and TolerantECG were rerun in the authors’ evaluation pipeline.The reruns serve as checks that the pipeline reproduces published values before comparing with numbers not generated by the authors.Source: §B.3 · p. 17
• Each main-result baseline was pretrained under both a shared search protocol and its paper’s specified schedule, with the configuration selected by validation loss.Objective-specific hyperparameters remained at published values in both runs.Source: §B.3 · p. 17
• ZZU-pECG includes 12,334 twelve-lead and 1,856 nine-lead records with variable record lengths.The appendix notes that external dataset handling and benchmark comparisons follow dataset-specific protocols.Source: §B.3 · p. 17, Table 12 · p. 18

C.1 ZZU-PECG

• ZZU-pECG contains 14,190 hospital ECG recordings from 11,643 children aged 0–14, with heterogeneous lead sets and record lengths from 5 to 120 seconds.Long records are split into non-overlapping 30-second windows, whose logits are averaged by record; macro AUROC is reported under a patient-level partition.Source: §C.1 · p. 18
• ZZU-pECG tests transfer across both recording setting and developmental age: it is hospital diagnostic ECG and spans ages beyond the infant pretraining window.The paper describes this as the stricter transfer test for potential population memorization.Source: §C.1 · p. 18

C.2 PTB-XL

• PTB-XL uses 10-second twelve-lead records and a recommended stratified ten-fold split; evaluation covers five diagnostic superclasses under linear evaluation and fine-tuning.The shared single-channel encoder processes each lead independently, with concatenated embeddings and no learned cross-lead interaction; the twelve-versus-one-lead comparison therefore reflects additional independent evidence.Source: §C.2 · pp. 18–19

C.3 MIMIC-IV-ECG

• BeatGraph-Adult is pretrained on roughly 800,000 MIMIC-IV-ECG records, using about 2,200 hours and one lead to match the infant run’s signal budget and single-channel regime.PTB-XL is held out, and the adult and evaluation corpora come from different institutions; the pretraining recipe is otherwise identical to the infant run.Source: §C.3 · p. 19

D.1 SHARED PROTOCOL

• Fine-tuning hyperparameters were fixed a priori, with five seeds per fold and between-infant standard deviations; baseline pretraining used 100 epochs and also tested published settings.Each baseline received 30 seconds resampled to its published input rate and context length; shorter-context predictions averaged segment logits.Source: §D.1 · p. 19

D.2 ECG-FM

• ECG-FM retained temporal CMSC on six consecutive 5-second segments but removed Random Lead Masking and compressed pretraining from 240 to 100 epochs.Its published configuration uses 5-second segments and a 90.9 M-parameter encoder.Source: §D.2 · pp. 19–20, Table 13 · p. 20

D.3 HUBERT-ECG

• HuBERT-ECG’s single-channel input would yield 47 rather than 93 tokens, so the convolutional stride was halved to restore approximately 94 tokens.The adaptation retained 33% masking and changed positional encodings to plain temporal positions.Source: §D.3 · p. 20

D.4 MTAE

• MTAE used a one-channel ViT-B backbone, matching the published PTB-XL comparison setup, with a 75% masking ratio.Its decoder follows the published asymmetric design.Source: §D.4 · p. 21

D.5 SIMCLR AND BYOL

• SimCLR and BYOL shared a 1-D convolutional encoder and augmentation family, isolating their contrastive and self-distillation objectives.The methods used NT-Xent with negatives for SimCLR and a predictor with symmetric cosine loss and no negatives for BYOL.Source: §D.5 · p. 21

D.6 ADULT-PRETRAINED CHECKPOINTS

• Adult-pretrained checkpoints were applied to infant tasks without encoder retraining, confounding pretraining domain with single-channel lead adaptation.The single-channel signal was placed in Lead I and the remaining eleven leads were zero-filled.Source: §D.6 · p. 21

E LEAVE-ONE-INFANT-OUT EVALUATION

• Each leave-one-infant-out fold kept training, validation, and test infants disjoint; this appendix selected models by validation loss, unlike the main paper’s fixed protocol.The appendix results therefore use a different model-selection protocol.Source: E LEAVE-ONE-INFANT-OUT EVALUATION · p. 21, §E.1 · p. 21, Table 14 · p. 21

E.2 CLASS COUNTS PER INFANT

• Sleep fractions ranged from 28% to 65% across infants, and Table 16 reports each held-out infant’s results as mean ± standard deviation over three seeds.Sleep/wake and infant-state tasks share 4,712 windows; activity-source windows come from a separate, smaller annotation stream.Source: §E.2 · p. 22, §E.3 · p. 22, Table 15 · p. 22

E.4 INTERVAL ESTIMATES UNDER TWO UNITS OF ANALYSIS

• With four infants, the Student-t interval has three degrees of freedom; cluster-resampled session-level intervals are narrower, and both approaches agree on ordering and overlap.The infant is the appropriate unit because sessions from the same infant are not independent.Source: §E.4 · p. 22, Table 17 · p. 23

E.6 EFFECT OF VALIDATION-SET CONSTRUCTION

• Selecting on a validation infant increased sleep and infant-state performance on the infant4 test fold, while activity-source performance was unaffected.The two model-selection protocols can differ by a small margin.Source: Table 18 · p. 23, §E.6 · p. 23

E.5 STATISTICAL COMPARISON BETWEEN MODELS

• With n=4, the smallest attainable two-sided Wilcoxon p-value is 0.125, so the authors do not claim infant-level significance.They report effect sizes and intervals, treating differences as meaningful only when exceeding between-infant standard deviation and sharing a sign across all four folds.Source: §E.5 · p. 23

F.1 SIGNAL-QUALITY FILTERING AND PRETRAINING WINDOW COUNT

• Recording-level quality control removed recordings with fewer than five detected peaks, leaving 403,320 pretraining windows; artifact-containing windows were retained.No per-window quality index or artifact rejection was applied, trading a noisier objective for a distribution representative of naturalistic sessions.Source: §F.1 · p. 23

F.2 WINDOW LABELLING: MIXED LABELS AND TRANSITIONS

• Downstream windows were formed within single-label segments, discarding residuals under 30 seconds; activity-source windows had no transition filter.This construction prevents mixed-label windows, and all qualifying non-overlapping windows were retained for activity source.Source: §F.2 · pp. 23–24

F.3 CLASS BALANCING

• No class balancing was used; macro-averaging gives rare classes such as crying and quiet equal contribution to reported scores..Source: §F.3 · p. 24

F.4 SELF-SUPERVISED PROJECTION AND PREDICTION HEADS

• BYOL-style heads use a projector and online predictor, with an exponential-moving-average target encoder and projector; these and the masked-prediction decoder are discarded after pretraining.The target copies use decay 0.99 and have gradients disabled.Source: §F.4 · p. 24

F.5 OPTIMISATION, BATCH SIZE AND HARDWARE

• Pretraining used AdamW for 100 epochs, while downstream fine-tuning used differential learning rates for the task head and encoder-readout.Runs used an NVIDIA A100 or A40; one pretraining run took approximately 22–23 hours.Source: §F.5 · p. 24

G SIMPLE PHYSIOLOGICAL BASELINES

• Physiological baselines isolate interbeat-interval structure and beat morphology, ranging from mean RR logistic regression to 14 HRV features and boosted trees.The mean-RR baseline measures how much each task is explained by average rate alone.Source: G SIMPLE PHYSIOLOGICAL BASELINES · p. 24, §G.1 · pp. 24–25

G.1 BASELINE DEFINITIONS

• The HRV boosted-tree baseline captures nonlinear feature interactions; RR-only BeatGraph removes beat-waveform descriptors, while the from-scratch model isolates pretraining’s contribution.The RR-only model otherwise keeps the full architecture and training setup.Source: §G.1 · p. 25

H.1 EFFECT OF BYOL LOSS

• Removing BYOL costs 0.08 to 0.17 macro-F1 across the four tasks, and infant-state performance falls to 0.42 versus 0.49 for the same architecture trained from scratch.The infant-state comparison is described as the task with the finest label granularity.Source: §H.1 · p. 25
• Macro-F1 varies non-monotonically with BYOL-loss weight, peaking at 0.5 and falling at 1.0 on every task; the study uses 0.5 thereafter.Table 19 compares macro-F1 for physiological baselines and BeatGraph on the fixed subject-disjoint split.Source: §H.1 · p. 25, §G.2 · p. 25
• At BYOL weight zero, the infant-state score is below the from-scratch result, though the authors treat the 0.04 gap as suggestive given four folds.The masked-prediction-only objective can map every beat to a constant because its target is a stop-gradient pass over the same weights.Source: §H.1 · p. 25