Self-Supervised Pre-Training for Non-Insvasive Fetal Electrocardiography from Limited Open-Access Data: A Cross-Dataset QRS-Transfer Benchmark

Self-supervised pre-training has transformed adult electrocardiogram (ECG) analysis, but it is unclear whether it can help non-invasive fetal ECG (fECG), where open annotated corpora are orders of magnitude smaller. Self-supervised models have been applied to pregnancy ECG for prenatal stress detection on private cohorts. We ask a different question: on public abdominal recordings, which self-supervised objective, if any, transfers to held-out fetal QRS detection? We compare contrastive (SimCLR), variance-covariance (VICReg) and masked signal modelling objectives on one encoder, pre-trained on 68 open recordings with PhysioNet/CinC 2013 held out. Naive contrastive pre-training collapses (effective rank 2.5 of 64) and underperforms a random encoder. Per-window instance normalisation prevents collapse (effective rank 9.6). All three objectives beat random initialisation on held-out CinC 2013 (record-grouped cross-validation, $\pm$50 ms): masked modelling reaches $F_{1}=0.33{}\pm 0.13{}$ versus $0.27{}\pm 0.09{}$ (paired Wilcoxon $p=1.4\times 10^{-7}{}$; Cohen's $d_{z}=0.75$, 95% CI [0.039,0.072]). The numeric edge over SimCLR ($F_{1}=0.32{}$) is small, not significant ($p=0.14$), and shrinks to 0.004 across three training seeds. With $5\%$ of labels the frozen masked encoder matches a random encoder trained on $100\%$. Under full fine-tuning the gap vanishes ($F_{1}=0.50{}$ vs $0.50{}$, $p=0.53{}$); with 3-6 labelled records pre-training helps under the same 12-epoch protocol ($+0.02$-0.03) but not when fine-tuning runs longer. On the same records, supervised delineation networks reach $F_{1}=0.92$ and classical extractors 0.68, so the probe measures representation quality. Embeddings do not encode gestational age and remain dataset-separable. Code, configurations and seeds are released.