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

What it claims

Making the heartbeat the token, instead of a fixed-length patch, yields an ECG encoder that works on single-channel infant wearable data. On four infant tasks it reaches F1 between 0.589 and 0.923. The paper also gives three results the summary does not:

• PTB-XL, linear evaluation: BeatGraph matches ST-MEM but trails it after full fine-tuning.
• ZZU-pECG: 0.892 AUROC, within 0.001 of the best published self-supervised ECG model, not ahead of it.
• PTB-XL, adult pretraining, single channel: the same architecture pretrained on adult ECG instead scores 0.825 against 0.804 for ST-MEM under linear evaluation.

The released corpus is 3,361 unlabeled hours plus only 47.5 labeled hours (state and affect), from 143 infants aged 3–11 months, with subject-level splits.

How it works

1. Beat graph. R-peaks are found with NeuroKit using an infant-tuned configuration on 30 s windows sampled at 1000 Hz. Each beat is cropped around its R-peak, with pre/post lengths scaled from the window's median RR interval and clipped to fixed bounds. Crops run 350–675 samples (median 384, at a mean of 137 bpm). Each window keeps at most N=80 beats, zero-padded with a validity mask. The graph is complete over valid beats.
2. Beat encoder (shared). A 1-D conv stack (1→32→64 channels) followed by global average pooling. The result is concatenated with [RR_prev, RR_next], projected to d=128 and layer-normalised.
3. Temporal Transformer. 2 layers, 4 heads, pre-norm, sinusoidal positional encoding.
4. Graph attention. 3 residual GATv2 layers, 4 heads of width 32, attending over all beat pairs.
5. Readout. An attention-weighted mean concatenated with a layer-normalised per-dimension standard deviation over beats, giving z ∈ R^256. The std term is there to preserve heart-rate variability, which a plain mean discards.

flowchart LR
  W[30s window] --> M[Mask 30% of beats]
  W --> T[Unmasked pass, stop-grad]
  M --> E[Transformer + GATv2] --> D[Linear decoder]
  D --> L[L2 on masked beats]
  T --> L
  W --> B[Two augmented views, EMA target]
  B --> Y[BYOL cosine loss x0.5]
  L --> O[Total loss]
  Y --> O

Pretraining has two terms:

• Masked beat-embedding prediction. 30% of valid beats are replaced by a learned mask token before the Transformer. The model predicts the post-Transformer embeddings H(1) of an unmasked, stop-gradient pass.
• BYOL regulariser. λ=0.5, EMA decay 0.99. It is there to block constant-embedding collapse.

The schedule is 100 epochs, using the final checkpoint.

Fine-tuning uses a residual-MLP head for 20 epochs, again using the final checkpoint. For 12-lead external data, the single-channel beat encoder runs per lead, the 12 embeddings are concatenated, and missing leads are zero-filled. Shorter external records form one graph each; the mask absorbs the smaller beat count.

Evidence

• Infant tasks: macro-F1 gains of 0.076–0.158 over the strongest baseline per task. The tasks are sleep-wake (2-class), infant state (4-class), activity source (2-class) and affect (2-class). Baselines are ECG-FM, HuBERT-ECG, BYOL, MTAE and ST-MEM. Per-task numbers are not in the fetched text.
• ZZU-pECG: 14,190 records from hospitalized children aged 0–14, 0.892 AUROC.
• PTB-XL: results as above, under both infant and adult pretraining.
• Ablations: listed as covering beat detection, graph attention, tokenization and beat-window capacity, but the results are not in the fetched text.

Limits

• High heart rates degrade the representation:
• 1.7% of windows exceed 160 bpm and are truncated to the first 80 beats.
• Above 171 bpm, adjacent crops overlap (5.1% of beat pairs).
• Depends on R-peak detection. The input is built from detected peaks, so detection errors on noisy wearable signals feed straight into the graph. The detection ablation's result is not shown.
• Small labeled set. The labels cover 47.5 h; the fine-tuning tasks rest on that.
• Weaker on adult ECG after full fine-tuning. For adult 12-lead use, infant pretraining loses to ST-MEM once fully fine-tuned.
• Status. The paper is marked "Submitted for review."
• Release details unknown. The fetched text is truncated mid-method. It gives no code or weights link, no corpus access procedure and no licence; check these before planning to use the data.