acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

Electrocardiogram (ECG) foundation models typically tokenize the signal into fixed-length patches that ignore cardiac structure, so a patch may split a heartbeat and the number of beats contained in each patch shifts as heart rate varies. This matters most for infants, whose heart rates are higher than adults and whose ECG differs from the adult, clinic-recorded 12-lead data on which these models are built. A model for infant ECG should therefore reason about heartbeats directly rather than recover them from arbitrary patches. We propose BeatGraph, which makes the heartbeat its unit of representation, modeling each 30-second window as a graph of beats. A shared beat encoder turns each heartbeat into a feature vector from its waveform and inter-beat intervals, a Transformer with positional encoding orders the beats in time, and residual graph attention layers relate every beat to every other before attention-based pooling produces a window-level representation. We pretrain BeatGraph on our new corpus of unlabeled infant recordings by predicting the embeddings of masked beats, then fine-tune it for each task. This single backbone supports sleep-wake detection, infant-state classification, activity-source identification, meaning whether movement is infant- or caregiver-initiated, and infant affect recognition, improving macro-F1 over the strongest baseline on each task by 0.076 to 0.158. It also transfers across age groups, reaching 0.892 AUROC on the ZZU-pECG pediatric benchmark among children 0 to 14 years of age, within 0.001 of the best published self-supervised ECG model, and matching that model under linear evaluation on the adult PTB-XL benchmark despite pretraining only on infant recordings. Finally, to our knowledge, we release the first public infant ECG corpus collected in homes, classrooms, and laboratory settings with state and affect labels. This corpus contains 3,408 hours of single-channel ECG from 143 infants aged 3 to 11 months, with unlabeled pretraining data, benchmark tasks, and subject-level splits.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
ECG representation learning, physiological time series, heartbeat tokenization, graph attention networks, masked prediction, wearable sensing, self-supervised learning

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_UkBtOpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
2
Presentation
1(−1)
Soundness
1
Read review
Paper summary

This paper investigates how fixed length patching may disrupt the structure of ECG heartbeats, with a focus on infant ECG, where heart rates are higher and recording conditions are more challenging. The authors propose BeatGraph, which represents each heartbeat as a graph node and combines waveform morphology and interbeat intervals to learn representations using a Transformer and graph attention. Experiments show that BeatGraph outperforms baseline models on four infant state recognition tasks and demonstrates some transferability to pediatric and adult ECG datasets. The authors also introduce a dataset containing 3408 hours of ECG recordings from 143 infants.

Critical strengths

The research topic is valuable, and infant ECG deserves further investigation. The authors collect long-term infant ECG recordings in natural environments, addressing the limited availability of public data in this area. The model achieves clear improvements across four infant tasks and performs well in cross-age transfer experiments. These results support the potential of heartbeat-level representations for infant ECG modeling.

Fixed length patching may split individual heartbeats across multiple tokens, making it harder for models to learn waveform morphology and rhythm. Previous studies have provided some evidence of this issue. BeatGraph extends this idea to infant ECG, and Table 7 further shows that heartbeat alignment improves downstream performance, providing direct evidence for the design choice.

Critical weaknesses

Motivation. Heartbeat-level tokenization has already been extensively studied. HeartLang [1] proposes using heartbeats as basic representation units, while two studies in 2026 [2, 3] directly examine and validate this approach. Therefore, the central motivation of BeatGraph lacks novelty. The paper also fails to demonstrate that higher heart rates make fixed length patching more problematic for infant ECG.

Model. The model introduces a fully connected GAT after the Transformer, although both components can capture global relationships between heartbeats. While the ablation study shows some benefits from the GAT, the paper does not clearly explain what the graph structure uniquely contributes. This makes the main methodological contribution appear closer to a combination of existing techniques.

Data. The labeled data for three tasks, including sleep, infant state, and activity source recognition, come from only four infants. Although leave-one-subject-out validation is used, the small number of subjects limits the reliability of the results. Furthermore, the emotion recognition task uses experimental phases as emotion labels. The model may therefore be learning differences between experimental conditions rather than the infants' actual emotional states. These issues weaken the claims about the model's applicability in real-world settings.

[1] Reading Your Heart: Learning ECG Words and Sentences via Pre-training ECG Language Model, ICLR 2025.

[2] Beat-Synchronous Tokenization for ECG Transformers, arXiv 2026.

[3] On the Role of the Tokenizer in ECG Transformer Models, arXiv 2026.

Although arXiv papers have not necessarily undergone peer review and should be treated with caution, HeartLang explicitly points out that fixed length patching ignores ECG waveform morphology and rhythm structure, potentially affecting representation learning.

Questions for authors
  1. Distinction from prior work. Methods such as HeartLang have already adopted heartbeat-level representations. What is the core novelty of BeatGraph compared with these methods? Does BeatGraph still offer an advantage when evaluated on the same infant data with comparable model capacity?

  2. Necessity of the graph structure. The Transformer can already model relationships between arbitrary heartbeats. What additional capability does the fully connected GAT provide? Can the authors demonstrate its necessity through more rigorous capacity-matched experiments or an analysis of its underlying mechanism?

  3. Generalization and label validity. Only four infants have annotations for three core tasks. Is validation on more independent subjects available? For emotion recognition, can the authors demonstrate that the model learns genuine emotional information rather than differences between experimental phases?

The paper's current presentation is rather weak, with an unconvincing motivation and insufficiently informative experimental results. The authors need to develop a clearer and more coherent narrative and address the concerns raised above. I find the research topic interesting, and I would be willing to raise my score if the authors can address these issues convincingly.

10 comments
AuthorsOpenReview ↗

Response to reviewer UkBt (1/8)

W1 / Q1. Motivation and distinction from prior work

We thank the reviewer for asking us to clarify the distinction from prior heartbeat-based methods and substantiate the motivation concerning infant heart rates. Heartbeat segmentation provides a common starting point; the contribution is how these segments are represented, related, and pretrained for infant wearable ECG. BeatGraph’s contribution is a compact representation-learning framework that combines continuous beat representations, explicit inter-beat timing, graph-based modeling, and masked self-distillation for infant wearable ECG. We have clarified this distinction and added comparisons with HeartLang and a capacity-matched fixed-patch baseline.

1. How does BeatGraph differ from HeartLang?

HeartLang is the closest prior work, and we now explicitly discuss and compare against it. Although both methods begin with heartbeat detection, they use the resulting segments differently. HeartLang converts beats into discrete words through a learned codebook and predicts the codebook indices of masked words. BeatGraph represents each beat as a continuous node, incorporates adjacent RR intervals as node features, models relationships between beats through graph attention, and learns by regressing masked-beat embeddings with self-distillation. The distinction therefore extends beyond segmentation to the representation, relational encoder, and learning objective.

Table R1. Comparison with HeartLang (Jin et al., 2025).

ComponentHeartLangBeatGraph
Beat detectionWavelet-based QRS detection on lead I; boundaries shared across leadsInfant-tuned NeuroKit configuration on the single channel (Table 9)
Beat segmentationMidpoint-to-midpoint segments, zero-padded to 96 samplesAdaptive asymmetric crop based on the window’s median RR interval: 0.38 before and 0.52 after the R-peak; no within-beat padding (Eq. 1)
Sampling rate100 Hz1000 Hz
Beat representationDiscrete codebook with 8,192 entriesContinuous, average-pooled convolutional descriptor
Rhythm informationImplicit in segment length, with additional per-second temporal embeddingsPrevious and next RR intervals explicitly included in each node (Eq. 2)
Inter-beat modelingSelf-attention over a flat sequence of beat words from all leadsTemporal Transformer followed by graph attention over beats
Lead handlingTwelve leads combined with lead embeddingsShared single-channel encoder applied independently to each lead
Window readout[CLS] tokenAttention-weighted mean with a dispersion term (Eq. 6)
PretrainingCodebook learning followed by masked code predictionSingle-stage masked continuous-embedding regression with self-distillation (Eq. 7)
Encoder parameters44.5 M1.37 M
Original pretraining settingAdult clinical ECG, 12 leads, 10 sInfant wearable ECG, single channel, 30 s
Evaluation focusAdult diagnostic classificationInfant behavioral tasks, with pediatric and adult transfer

Component ablations further support these design choices: the infant-tuned detector in Table 9, explicit inter-beat intervals in Table 7, and the self-distillation term in Table 20.

These choices are particularly relevant to infant ECG. HeartLang pads each beat word to 96 samples at 100 Hz, corresponding to 960 ms. At the infant mean RR interval of 438 ms, a typical midpoint-to-midpoint segment occupies approximately 44 samples, leaving more than half of the word as padding. A QRS complex lasting approximately 60 ms is represented by only six samples at this sampling rate. These observations motivate an adaptive representation that preserves higher-resolution morphology and explicitly encodes timing. HeartLang also identifies limitations involving zero padding under baseline drift and irregular QRS complexes in Appendix B.3, which are relevant to noisy wearable recordings.

The learning objectives differ as well. HeartLang reports that replacing its vocabulary-based target with direct reconstruction is unsuccessful in its setting (Table 3). BeatGraph instead predicts continuous latent embeddings with self-distillation, whose contribution is examined in our Table 20. Our results therefore establish the effectiveness of a different continuous-target formulation that eliminates the separate codebook-learning stage.

AuthorsOpenReview ↗

Response to reviewer UkBt (2/8)

W1 / Q1. Motivation and distinction from prior work (Continued)

2. Relationship to the two recent tokenization studies

We now cite both [2] and [3], posted on August 31 and September 14, 2026, respectively and consider as our concurrent work. These studies provide useful context for heartbeat-aligned tokenization but address a different methodological question. Study [2] compares tokenizers on adult PTB-XL and Icentia11k and finds beat-aligned tokens comparable to fixed patches. Study [3] examines tokenization as the variable in a supervised study on CPSC2018. Neither investigates infant ECG or introduces the continuous beat-graph encoder and self-distillation framework studied here.

3. Comparisons of BeatGraph with HeartLang, CMSC, and ST-MEM

To examine whether BeatGraph’s gains can be explained by model capacity or access to infant pretraining data, we pretrained a fixed-patch encoder with 30 ms patches and closely matched capacity: 1.40 M parameters versus BeatGraph’s 1.37 M. Both used the same infant corpus and pretraining budget. We additionally pretrained HeartLang, CMSC, and ST-MEM on this corpus. We also report the fine-tuned ST-MEM results from Table 5.

Table R2. Macro-F1 on four infant tasks. BeatGraph’s first three results are from Table 1; affect is evaluated on the released split over five seeds. Fixed-patch, CMSC, and infant-pretrained ST-MEM results use 12 runs for the classroom tasks (four test infants × three seeds) and five seeds for affect. HeartLang uses four test infants × five seeds for the classroom tasks and five seeds for affect. The final row reproduces the fine-tuned ST-MEM results from Table 5.

ModelEncoder parametersSleep/wakeInfant stateActivity sourceAffect
BeatGraph1.37 M0.9230.589<u>0.745</u>0.731
Fixed 30 ms patches1.40 M0.7760.4780.6290.593
CMSC (CLOCS)0.05 M0.7440.3230.5160.603
HeartLang44.5 M0.8370.4910.5970.557
ST-MEM, infant-pretrained85.1 M0.8650.5710.7490.637
ST-MEM, fine-tuned (Table 5)—0.68 ± 0.060.39 ± 0.040.52 ± 0.020.52 ± 0.04

These comparisons support four observations.

  1. The advantage persists at matched capacity. BeatGraph exceeds the fixed-patch encoder by 0.187, 0.111, 0.216, and 0.178 macro-F1 on sleep/wake, infant state, activity source, and affect, respectively. Matching model size, pretraining data, and budget rules out these factors as explanations for the gains. This comparison supports BeatGraph’s overall beat-centered design, while the component ablations examine individual design choices.

  2. Beat alignment alone does not account for the results. BeatGraph outperforms HeartLang on all four tasks by 0.086, 0.098, 0.148, and 0.174 macro-F1, despite using approximately 32 times fewer encoder parameters.

  3. BeatGraph remains competitive with a substantially larger infant-pretrained encoder. With approximately 62 times fewer parameters, BeatGraph exceeds infant-pretrained ST-MEM on sleep/wake, infant state, and affect by 0.058, 0.018, and 0.094 macro-F1, respectively, while trailing by only 0.004 on activity source. This supports a favorable performance–capacity tradeoff.

  4. The fine-tuned ST-MEM results provide an additional comparison. Relative to the reported means in Table 5, BeatGraph achieves higher macro-F1 on all four tasks, with absolute differences of approximately 0.24, 0.20, 0.23, and 0.21. These results complement the controlled pretraining comparisons above.

Together, the architecture comparison, aggregate metrics, and individual-infant results provide complementary evidence for the benefit of BeatGraph’s continuous beat representations and relational learning framework.

AuthorsOpenReview ↗

Response to reviewer UkBt (3/8)

W1 / Q1. Motivation and distinction from prior work (Continued)

4. Does higher heart rate strengthen the case for beat-centered representations?

We now substantiate this motivation with complementary evidence from patch alignment, the matched-capacity comparison, and heart-rate-stratified evaluation.

Patch alignment. Median heart rate across the evaluated tasks is 123–132 bpm, corresponding to approximately 455–488 ms between beats. At these rates, moderately long fixed patches can contain multiple R-peaks, while short patches divide QRS morphology across token boundaries. Table R3 quantifies this tradeoff in the infant test recordings.

Table R3. Alignment of fixed patches with detected beats in the sleep and state test windows, whose median heart rate is 123 bpm. QRS overlap is assessed using the interval R ± 40 ms.

Patch length (ms)Beats with a patch boundary inside R ± 40 msPatches with no R-peakPatches with ≥2 R-peaks
30100.0%94.0%0.0%
25032.1%50.3%0.0%
50015.7%12.9%12.3%
65012.2%9.9%39.1%
10007.7%6.5%82.7%

These statistics do not imply that fixed-patch models cannot learn across boundaries. They demonstrate the segmentation tradeoff motivating beat alignment: shorter patches fragment local morphology, whereas longer patches increasingly combine multiple beats. BeatGraph directly associates each node with a detected beat and separately represents its timing.

Matched capacity. This motivation is accompanied by an empirical advantage: BeatGraph improves macro-F1 by 0.111–0.216 on the three classroom tasks relative to the capacity-matched 30 ms fixed-patch encoder.

Heart-rate-stratified evaluation. We split each task’s test windows at its median heart rate and evaluated both models on identical subsets in paired runs.

Table R4. Macro-F1 in the slower and faster halves of each task. Results are mean ± standard deviation over three seeds, with the four test infants pooled within each run.

TaskMedian HR (bpm)BeatGraph, slowerBeatGraph, fasterFixed-patch, slowerFixed-patch, fasterGap, slowerGap, faster
Sleep/wake1230.913 ± 0.0580.841 ± 0.0090.805 ± 0.0560.709 ± 0.0170.1080.132
Infant state1230.512 ± 0.0380.590 ± 0.0130.442 ± 0.0100.450 ± 0.0160.0700.140
Activity source1320.795 ± 0.0290.812 ± 0.0130.698 ± 0.0230.619 ± 0.0720.0970.193

BeatGraph’s advantage increases in the faster-heart-rate subset for all three tasks, doubling for infant state and approximately doubling for activity source. This pattern is consistent with the proposed motivation. Because heart rate can also covary with behavioral state and signal quality, we interpret this analysis as supporting evidence rather than causal proof that heart rate alone determines the performance gap.

Reviewer_Xxb1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors propose a beat-based model for single-lead infant ECG classification. The model consists of a 1D convolutional encoder with pooling, a Transformer, and a graph attention network, followed by attention-weighted pooling that preserves beat-to-beat variability within a 30-second window. They pretrain the model in a self-supervised manner on their collected infant ECG data by predicting masked beat embeddings, with a BYOL regularizer.

They plan to release an infant ECG dataset collected in naturalistic, nonclinical environments. The dataset contains 3,408 hours of single-lead ECG recordings from 143 infants aged 3 to 11 months, including unlabeled data and labels for four classification tasks. Three of these tasks are labeled using data from 4 infants. The paper also evaluates an adult ECG dataset PTB-XL.

Their method performs well on the infant tasks and transfers reasonably well to adult multi-lead ECG, especially in the linear-evaluation setting. Overall, the work contributes a valid single-lead ECG model and evaluates it on four infant ecg classification tasks.

Critical strengths
  1. The experiments on infant data demonstrate the effectiveness of beat-based modeling compared with window- or waveform-based modeling, as shown in Table 1. The proposed method can effectively learn representations from single-lead ECG, and it generalizes well on multi-lead PTB-XL dataset.
  2. The authors introduce a non-clinically collected infant ECG dataset comprising 3,408 hours of recordings from 143 infants, which could be a meaningful resource for infant ECG research.
  3. The paper is of good quality from an experimental perspective and makes meaningful contributions to the particular application area.
Critical weaknesses
  1. Insufficient novelty. The method design and self-supervised training are not sufficiently novel. The technical gap between adult and infant ECG is not clearly enough except differences in the number of leads and heart rate. In other words, there is insufficient evidence that a multi-lead method, after being adapted to single-lead ECG and self-supervisedly pretrained on the same single-lead infant data, would fail. Beat tokenization is the strongest insight of the framework, but beat-based ECG modeling is not new. The paper therefore needs to clarify whether its main contribution is intended to be a new architecture and corresponding self-supervised learning, or a new infant ECG dataset and benchmark.

Table 1 should include stronger adapted multi-lead baselines, or other single-lead baselines. For example, the authors could adapt ST-MEM(cited), or beat-based baselines, such as BLF-Net (Jing et al., 2023), beat-based self-supervised leanring, e.g., Beat-SSL(2026), to single-lead and pretrain them on the same infant corpus.

  1. Insufficient analysis of the Transformer–GAT design. The practical, or physiological roles of the Transformer and GAT are insufficiently investigated. The graph is fully connected, meaning that both the Transformer and GAT model beat-to-beat interactions. The paper does not provide an in-depth analysis of what the GAT captures beyond self-attention.

  2. From a benchmark perspective, three of the four downstream tasks involve only four infants and 2–4 classes. The challenge and generalizability may be insufficient to support broader research.

Questions for authors
  1. When fine-tuning ST-MEM to infant data, the authors zero-pad the remaining leads and place the single lead in Lead I. I am not sure whether this experimental design is sufficient. For example, could you try non-zero padding, such as replicating the single lead across all 12 leads, or removing/adapting the lead-specific modules?
  2. Table 5 evaluates only the direct transfer of off-the-shelf adult multi-lead models. It does not evaluate whether their architectures could be adapted to single-lead input and pretrained on the same unlabeled infant corpus. For example, considering adding ST-MEM and CMSC/Beat-SSL to Table 1.
  3. What is the difference between the Transformer layers and graph-attention layers when using a fully connected graph, especially given that 32-kNN performs competitively? For example, it might be worthy to investigate whether the GAT exploits any meaningful physiological topology.
  4. After global pooling over the temporal dimension following the convolutional encoder, within-beat temporal information is compressed. All subsequent modules model interactions among heartbeats within a window. How do you interpret the physiological information retained in this pooled beat representation? Could information lost through temporal pooling be important for downstream tasks?
  5. The practical and clinical value of non-clinical wearable ECG collected in naturalistic settings also needs clearer articulation.
  6. The weight in ‘Readout’ of Figure 2 is learned or from the GAT? It is not stated in 3.2.
11 comments
AuthorsOpenReview ↗

Response to reviewer Xxb1 (1/9)

W1. Methodological contribution and distinction from prior work

We thank the reviewer for asking us to clarify the contribution and strengthen the baseline comparisons. BeatGraph contributes a compact self-supervised framework for infant wearable ECG, with a released corpus and benchmark that enable research in this setting. Its contribution lies in how continuous beat morphology, explicit rhythm, and learned inter-beat relations are represented and trained together. The additional experiments directly address whether the results can be explained by heartbeat segmentation, model capacity, or the adaptation of adult models.

The representation and learning objective distinguish BeatGraph from existing beat-based methods. HeartLang quantizes beats into a learned vocabulary and predicts masked codebook indices. BeatGraph instead retains continuous beat descriptors, incorporates the preceding and following RR intervals, applies GATv2 to contextualized beat representations, and regresses masked latent embeddings with self-distillation. This formulation avoids a separate codebook-training stage. The detector, RR-feature, and self-distillation ablations in Tables 9, 7, and 20 support the corresponding design choices; heartbeat segmentation itself is established practice.

The benefit is supported by controls on both capacity and architecture. At approximately matched encoder size and with the same infant corpus and pretraining budget, BeatGraph exceeds the fixed-patch comparator by 0.187, 0.111, 0.216, and 0.178 macro-F1 across sleep/wake, infant state, activity source, and affect-related episode discrimination. The architectural controls further show that adding three GATv2 layers improves the two-layer Transformer by 0.14, 0.13, and 0.17 on the classroom tasks with approximately 0.15 M additional parameters. Four additional Transformer layers use approximately 2.37 M additional parameters and yield smaller improvements on every task. The attention analysis and neighborhood controls in W2/Q3 explain the complementary role of relational refinement; the probes in Q4 show that the pooled beat descriptors retain information about morphology and within-beat timing.

Stronger adaptations and same-corpus pretraining reinforce the comparison. Q1 evaluates four ST-MEM input adaptations. Q2 adds infant-pretrained ST-MEM, CMSC, HeartLang, and Beat-SSL, including both discrete beat-token and contrastive alternatives. BeatGraph achieves higher reported macro-F1 than CMSC, HeartLang, and Beat-SSL on all four tasks, and the highest macro-F1 among these models on three tasks. It remains within 0.004 of infant-pretrained ST-MEM on activity source while using 1.37 M versus 85.1 M encoder parameters. These results establish a strong performance–size tradeoff against architectures trained directly on infant data.

The dataset contribution makes this methodological question practically consequential. Long naturalistic recordings, movement-related variability, and sparse behavioral labels create a different learning setting from short adult diagnostic ECG. The released infant recordings and evaluation tasks make this setting accessible for reproducible comparison, while 3,361 hours of unlabeled home ECG support self-supervised learning. The resulting contribution combines a physiologically motivated representation, experimentally supported architectural choices, and a reusable infant ECG resource.

AuthorsOpenReview ↗

Response to reviewer Xxb1 (2/9)

Q1. Alternatives to zero-padding for adult-pretrained ST-MEM

We evaluated the reviewer’s suggested alternatives using the same infant signal: zero-padding the other 11 leads; replicating the signal across all 12 leads; retaining a single lead with the Lead I embedding and separators; and using a single lead without lead embeddings or separators.

Preprocessing and augmentation precede lead arrangement, so all configurations receive the same underlying ECG and padded leads remain zero. The four configurations share 250 Hz input, 0.67–40 Hz band-pass filtering, random 9 s training crops, and predictions averaged over three crops per 30 s test window. Optimization, splits, and checkpoint selection are held fixed across configurations.

Table R1. Macro-F1 for adult-pretrained ST-MEM under four single-lead adaptations. Bold denotes the highest mean within each task.

TaskZero-paddingReplication across 12 leadsSingle lead with Lead I embedding and separatorsSingle lead without lead embedding or separators
Sleep/wake0.68 ± 0.060.73 ± 0.040.67 ± 0.040.71 ± 0.06
Infant state0.39 ± 0.040.43 ± 0.050.41 ± 0.050.44 ± 0.05
Activity source0.52 ± 0.020.57 ± 0.040.54 ± 0.030.55 ± 0.04
Affect-related episode discrimination0.52 ± 0.040.584 ± 0.0170.532 ± 0.0120.575 ± 0.014

The controls confirm that adaptation improves ST-MEM, while preserving BeatGraph’s reported advantage. Replication performs best on three tasks, and removing lead-specific modules performs best on infant state. Even choosing the strongest adaptation separately for each task gives macro-F1 of 0.730, 0.440, 0.570, and 0.584, compared with 0.923, 0.589, 0.745, and 0.731 for BeatGraph: differences of 0.193, 0.149, 0.175, and 0.147, respectively. Thus, the advantage is not eliminated by the alternative lead arrangements requested by the reviewer.

ST-MEM’s multi-lead pretraining exploits complementary spatial measurements that a single infant lead does not provide. Replication changes the input arrangement but cannot recreate those measurements. This provides a motivation for adapting and pretraining the architecture directly on infant ECG, which we evaluate separately in Q2; the lead controls do not isolate the decoder as the cause of the transfer gap. The affect adaptation runs use test-based checkpoint selection and are reported as exploratory comparisons under that shared protocol.

AuthorsOpenReview ↗

Response to reviewer Xxb1 (3/9)

Q2. Single-lead baselines pretrained on the same infant corpus

We have expanded the Table 1 comparison to include ST-MEM, CMSC (CLOCS), HeartLang, and Beat-SSL pretrained on the unlabeled infant recordings. This directly evaluates the reviewer’s proposed alternative: adapting existing architectures to single-lead input and learning from the target corpus. ST-MEM supplies a masked-autoencoding baseline, CMSC supplies patient-specific temporal contrastive learning, HeartLang supplies discrete beat-token modeling, and Beat-SSL supplies an additional beat-based self-supervised baseline.

Table R2. Infant-pretrained baseline comparison. Bold denotes the highest mean per task and metric; BeatGraph is italic when second-best.

TaskModelAccuracyMacro-F1Cohen’s κ
Sleep/wakeBeatGraph0.941 ± 0.0260.923 ± 0.0380.812 ± 0.057
Sleep/wakeST-MEM0.887 ± 0.0330.865 ± 0.0620.730 ± 0.053
Sleep/wakeCMSC0.783 ± 0.0650.744 ± 0.1050.612 ± 0.041
Sleep/wakeHeartLang0.855 ± 0.0290.837 ± 0.0300.676 ± 0.059
Sleep/wakeBeat-SSL0.857 ± 0.0720.815 ± 0.1230.638 ± 0.233
Infant stateBeatGraph0.732 ± 0.0230.589 ± 0.0270.483 ± 0.038
Infant stateST-MEM0.709 ± 0.0560.571 ± 0.0590.443 ± 0.058
Infant stateCMSC0.533 ± 0.0730.323 ± 0.0610.083 ± 0.072
Infant stateHeartLang0.640 ± 0.0510.491 ± 0.0520.435 ± 0.105
Infant stateBeat-SSL0.721 ± 0.0590.552 ± 0.0680.487 ± 0.129
Activity sourceBeatGraph0.767 ± 0.0150.745 ± 0.0210.513 ± 0.031
Activity sourceST-MEM0.781 ± 0.0350.749 ± 0.0390.527 ± 0.039
Activity sourceCMSC0.542 ± 0.0530.516 ± 0.0640.245 ± 0.061
Activity sourceHeartLang0.638 ± 0.1200.597 ± 0.1300.291 ± 0.198
Activity sourceBeat-SSL0.656 ± 0.1020.616 ± 0.1260.302 ± 0.197
Affect-related episode discriminationBeatGraph0.812 ± 0.0140.731 ± 0.0190.423 ± 0.023
Affect-related episode discriminationST-MEM0.747 ± 0.0190.637 ± 0.0090.295 ± 0.021
Affect-related episode discriminationCMSC0.641 ± 0.0230.603 ± 0.0170.186 ± 0.023
Affect-related episode discriminationHeartLang0.571 ± 0.0290.557 ± 0.0200.127 ± 0.039
Affect-related episode discriminationBeat-SSL0.573 ± 0.0100.544 ± 0.0160.133 ± 0.027

BeatGraph achieves the highest reported macro-F1 on three of four tasks with approximately 1/62 of ST-MEM’s encoder parameters. Relative to ST-MEM, its macro-F1 is higher by 0.058 on sleep/wake, 0.018 on infant state, and 0.094 on affect-related episode discrimination, and lower by 0.004 on activity source. ST-MEM uses approximately 85.1 M parameters, compared with 1.37 M for BeatGraph. BeatGraph also exceeds CMSC, HeartLang, and Beat-SSL in macro-F1 on every task, with gains of 0.128–0.266, 0.086–0.174, and 0.037–0.187, respectively.

Comments on the paper

5 comments
AuthorsOpenReview ↗

Response to reviewer kzzf (1/4)

W1 / Q1. Consistency across held-out infants and breadth of evaluation

We thank the reviewer for the careful distinction between subject-level generalization, the interpretation of episode labels, and the reliability of beat-based processing.

We have added three stronger infant-pretrained baselines: ST-MEM, HeartLang, and Beat-SSL. These comparisons include a large masked-autoencoding architecture and two beat-based self-supervised alternatives, all adapted to the infant recordings. Table R1 shows new baselines' performance, and then Table R2 shows the results for each held-out infant to examine whether the advantages are concentrated in particular folds.

Table R1. Aggregate macro-F1 for BeatGraph and the three added baselines. Bold denotes the highest mean; BeatGraph is in italics when second-best. ST-MEM has approximately 62 times as many encoder parameters as BeatGraph.

ModelEncoder parametersSleep/wakeInfant stateActivity sourceAffect-related episode discrimination
BeatGraph1.37 M0.9230.5890.7450.731
ST-MEM85.1 M0.8650.5710.7490.637
HeartLang44.5 M0.8370.4910.5970.557
Beat-SSL1.416 M0.8150.5520.6160.544

BeatGraph achieves the highest aggregate macro-F1 on three of four tasks and is within 0.004 of ST-MEM on activity source. It exceeds HeartLang and Beat-SSL on all four tasks, with improvements of 0.086–0.174 and 0.037–0.187, respectively. Relative to ST-MEM, the improvements are 0.058, 0.018, and 0.094 for sleep/wake, infant state, and affect-related episode discrimination, with a substantially smaller encoder.

The per-infant comparison shows that the gains extend across held-out infants. Table R2 adds the supplied ST-MEM and Beat-SSL fold results to the existing BeatGraph–HeartLang comparison.

Table R2. Macro-F1 paired by held-out infant. Bold denotes the highest mean within each infant–task row; BeatGraph is in italics when second-best.

TaskHeld-out infantBeatGraphST-MEMHeartLangBeat-SSL
Sleep/wakeInfant 10.9280.8880.8630.889
Sleep/wakeInfant 20.9330.8260.8330.896
Sleep/wakeInfant 30.9190.8440.8640.871
Sleep/wakeInfant 40.9310.9030.7890.603
Infant stateInfant 10.5850.5660.5600.642
Infant stateInfant 20.6330.5710.4270.591
Infant stateInfant 30.5930.5020.5190.474
Infant stateInfant 40.5890.6450.4580.502
Activity sourceInfant 10.7760.7790.5930.493
Activity sourceInfant 20.7210.7060.5200.647
Activity sourceInfant 30.7880.8190.8090.808
Activity sourceInfant 40.7200.6930.4660.515

BeatGraph exceeds HeartLang in 11 of 12, Beat-SSL in 10 of 12, and ST-MEM in 9 of 12 infant–task comparisons. It leads all three baselines on every infant for sleep/wake. For infant state, it exceeds HeartLang on all four infants and each of ST-MEM and Beat-SSL on three of four. For activity source, it exceeds HeartLang and Beat-SSL on three of four infants and ST-MEM on two of four.

The exceptions are visible in the table: Beat-SSL leads infant state for infant 1, ST-MEM leads infant state for infant 4, and ST-MEM leads activity source for infants 1 and 3. The overall pattern supports an advantage distributed across the cohort, while making the task- and infant-specific variation explicit.

The classroom evaluation provides both subject separation and longitudinal coverage. Its four infants are disjoint from the home-pretraining cohort. The recordings span 45 sessions over 7–15 weeks and 39.3 hours, including 4,712 windows for sleep/wake and infant state and 2,669 windows for activity source. Trained research assistants annotated behavior live, followed by video review and correction. This supports evaluation across repeated observations of each infant while retaining the infant as the independent unit.

The four-infant sample remains a limitation for broad population-level claims. In particular, an exact two-sided Wilcoxon signed-rank test with four nonzero pairs has a minimum attainable p-value of 0.125. We therefore emphasize the observed effect sizes and fold consistency rather than claim infant-level statistical significance. The observed consistency is complemented by the larger subject-disjoint adult evaluations, which test the representation in additional cohorts and recording settings.

AuthorsOpenReview ↗

Response to reviewer kzzf (2/4)

W2 / Q2. Pretraining exposure, episode labels, and the released dataset

The laboratory task evaluates affect-related episode discrimination with prior unlabeled exposure to the laboratory cohort. The 122 laboratory infants belong to the home-recording cohort used for self-supervised pretraining. The labeled laboratory recordings are excluded from pretraining, and the 24 laboratory test infants are excluded from supervised training and validation. However, these infants are not an entirely unseen cohort with respect to the unlabeled home data.

The available experiment does not separately evaluate previously exposed and entirely unseen laboratory infants. The shared home-recording cohort is the available source of infant pretraining data, and we do not have a separate evaluation excluding the laboratory test infants from pretraining. We retain this as a limitation of our work. In contrast, the classroom cohort is disjoint from the infant pretraining cohort, and the adult evaluations provide additional subject-disjoint downstream tests on external datasets.

The labels identify social-interaction episodes, and their physiological relevance is supported by prior developmental research. Play and still-face episodes define reproducible conditions in the Still-Face Paradigm. Ritz et al. studied 278 six-month-old infants and found associations between cardiac autonomic responses and independently observed affect, accounting for physical activity and respiration [1]. This motivates the use of ECG in the paradigm. It does not provide individual affect annotations for our recordings: our released labels remain episode identities, and we describe the task as affect-related episode discrimination, rather than independently validated emotion recognition.

Discrimination across more than one episode contrast. Starting with the released 98/24 infant split, we reserve 20 of the 98 training infants for validation, fit on the remaining 78, and retain the same 24 test infants.

Play versus still face achieves macro-F1 of 0.731 and AUROC of 0.748. Play versus reunion is harder, with macro-F1 of 0.670 and AUROC of 0.648. These results demonstrate that the learned representation supports discrimination across more than one social-interaction contrast.

Table R3. Affect classification on Play vs Reunion vs still face

TaskAccuracyMacro-F1Cohen’s κAUROCTest windows
Play vs still face0.748 ± 0.0070.731 ± 0.0050.423 ± 0.0110.748 ± 0.014171
Play vs reunion0.673 ± 0.0160.670 ± 0.0140.369 ± 0.0330.648 ± 0.024188
Play vs still face vs reunion0.565 ± 0.0130.544 ± 0.0140.426 ± 0.0200.656 ± 0.009263

The release contributes a reusable resource for an under-resourced infant sensing setting. It combines 3,361 hours of unlabeled home ECG from 139 infants, longitudinal classroom behavioral recordings, and 8.2 hours of laboratory recordings from 122 infant–caregiver dyads, together with raw and processed data, task labels, subject splits, and evaluation code. The value of the resource is that other researchers can reproduce the benchmark, compare representations, and extend the available annotations. Adult affective resources such as WESAD and CASE [2,3] offer complementary benchmarks, but do not replace infant recordings and developmental tasks.

AuthorsOpenReview ↗

Response to reviewer kzzf (3/4)

W2 / Q2. Pretraining exposure, episode labels, and the released dataset (continued)

Additional evidence: adult evaluations supporting the generalization and label discussion

BeatGraph’s performance extends beyond the four-infant classroom cohort and laboratory cohort. We evaluate WESAD, which includes wearable physiological recordings from 15 adults [2], and CASE, which includes physiological recordings and continuous affect annotations from 30 adults [3]. Our specific tasks use WESAD protocol labels and CASE video categories, as detailed below; the CASE experiments do not use its continuous subjective ratings as targets.

Table R4. Adult evaluation setup.

DatasetSubjectsEvaluation foldsTask classesLabel source30 s windows per class
WESAD15 adults15, leave-one-subject-outAmusement / stressProtocol phase: funny clips / Trier Social Stress Test180 / 326
WESAD15 adults15, leave-one-subject-outBaseline / stress / amusementProtocol phase581 / 326 / 180
CASE30 adults6, subject-levelAmusing / scaryVideo category330 / 300
CASE30 adults6, subject-levelAmusing / boring / relaxed / scaryVideo category330 / 240 / 240 / 300

Table R5. Adult macro-F1, reproduced from R1. Bold denotes the highest reported mean.

TaskBeatGraphST-MEMECG-FMCMSCBYOLSimCLRBeat-SSL
WESAD, amusement vs stress0.7530.7130.6130.7100.6570.6530.712
WESAD, three classes0.5130.5060.3830.4340.4120.4030.469
CASE, amusing vs scary0.6460.6010.5040.5120.5470.5410.598
CASE, four video categories0.4120.3590.2470.2590.3120.3060.365

BeatGraph achieves the highest reported macro-F1 on all four adult tasks, with advantages over the strongest comparator of 0.040, 0.007, 0.045, and 0.047, respectively. This strengthens the evidence that the representation is useful across datasets, subject groups, and acquisition settings, rather than only for the classroom cohort.

These adult evaluations complement the infant results without increasing the number of independently labeled classroom infants. Their protocol- and stimulus-defined targets also retain their stated interpretation. Together, the evaluations support consistent performance across the observed infant folds and competitive performance on external adult cohorts, alongside the explicit limits of the current infant sample, episode labels, and pretraining exposure.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.