Self-Supervised Representation Learning of Fetal Heart Rate Outperforms a Foundation Model in Identifying Fetal Acidosis

Johann Vargas-Calixto1, Masoud Nateghi2, Jay Ward3, Jim Robertson4, Yuri V. Sebastião5, Jackie K Patterson6, Jeffrey S. A. Stringer7, Reza Sameni8, Gari D. Clifford8
1Emory University, 2PhD student, 3MindChild Medical, 4MindChild, 5UNC Chapel Hill, 6UNC-CH, 7University of North Carolina School of Medicine, 8Emory University and Georgia Institute of Technology


Abstract

Background: Label scarcity in fetal monitoring is a severe limitation to training deep learning models for acidosis detection models; blood gas testing is not universal practice. Self-supervised learning (SSL) can leverage unlabeled recordings to learn signal representations without outcome labels.

Methods: We used the SimMTM SSL framework to pretrain an encoder using unlabeled fetal heart rate (FHR) recordings. The frozen encoder extracted 256-dimensional features from the last hour before delivery, which were classified by a BiLSTM. We also trained the same architecture without encoder pretraining. The end-to-end model was compared to the pretrained model and to Google's TimesFM 2.5, a 200M-parameter time series foundation model. We evaluated the models using a stratified 10X5-fold cross-validation on 1,217 labeled patients (59 acidosis cases). External validation was performed on the public CTU-UHB dataset (N = 552, 29 acidosis cases).

Results: The SSL-pretrained model achieved AUROC of μ=0.704 ± σ=0.054, significantly outperforming the end-to-end baseline (0.677 ± 0.069; p = 0.006, rank-biserial r = 0.473). TimesFM did not improve over the baseline (0.684 ± 0.066; p = 0.608). On CTU-UHB, the SSL model achieved the highest AUROC (0.743), followed by end-to-end (0.637) and TimesFM (0.600).

Conclusions: Domain-specific SSL pretraining on FHR data significantly improves fetal acidosis detection over supervised training alone, while a general-purpose foundation model provides no benefit. The superiority of SSL was maintained during external testing. Thus, it is necessary to leverage unlabeled FHR data to address label scarcity in fetal monitoring.