Key result
FHRFormer accurately reconstructs missing fetal heart rate signals and forecasts short-term traces.
Why the study?
Wearable fetal heart rate monitors are prone to signal dropout from maternal movement and sensor displacement, which complicates automated analysis. Conventional interpolation methods often fail to preserve important temporal and spectral characteristics of fetal heart rate signals.
Does the FHRFormer self-supervised masked transformer framework accurately reconstruct missing segments and forecast short-term dynamics in continuous fetal heart rate recordings during labor?
Does the FHRFormer self-supervised masked transformer framework accurately reconstruct missing segments and forecast short-term dynamics in continuous fetal heart rate recordings during labor?
FHRFormer provides a self-supervised deep learning method to reconstruct missing Doppler fetal heart rate data from wearable monitors and forecast short-term trends, potentially improving data completeness for future AI-based obstetric risk assessment.
Hypothesis-generating for transformer-based FHR reconstruction and forecasting; prospective outcome trials needed before clinical adoption.
Introduction Approximately 10% of newborns require assistance to initiate breathing at birth, and around 5% need ventilation support. Continuous fetal heart rate (FHR) monitoring supports the assessment of fetal well-being and timely obstetric interventions. However, wearable FHR monitors are prone to signal dropout caused by maternal movement and sensor displacement, resulting in missing data that complicate automated analysis. Conventional interpolation methods often fail to preserve important temporal and spectral characteristics of FHR signals. Methods We developed FHRFormer, a self-supervised masked transformer-based autoencoder for reconstructing missing FHR signals and forecasting short-term FHR dynamics. The model was trained on 4,486 FHR recordings collected using the Moyo monitor during labor in Tanzanian hospitals. By combining temporal reconstruction with frequency-aware loss functions, the framework aims to preserve both local signal morphology and spectral characteristics. Performance was evaluated using independent validation and test datasets, different masking ratios, and varying input patch sizes. Results FHRFormer demonstrated the ability to reconstruct missing FHR segments while preserving temporal and spectral characteristics consistent with surrounding signals. Reconstruction performance was highest with an input patch size of 30 samples and a masking ratio of 15%, with decreasing accuracy at higher masking ratios. The framework also demonstrated short-term FHR forecasting using a 30-minute historical context and a 15-second prediction horizon. Discussion FHRFormer offers a promising approach to improving the completeness of FHR recordings for retrospective research and AI-based analysis. Future integration into wearable FHR monitoring devices could support real-time risk detection. However, clinical benefits, including improved prediction of neonatal outcomes, remain to be established.
No takes yet. Share an insight, caveat, or question.
Engan et al. (2026) studied Fetal heart rate monitoring during labor (n=5,225). FHRFormer (self-supervised masked transformer-based autoencoder) vs. Conventional interpolation methods (e.g., linear interpolation) was evaluated on FHR signal reconstruction accuracy (evaluated via MSE, RMSE, MAE, PSNR, SSIM, FID, and CC) and short-term forecasting. FHRFormer reconstructed missing FHR segments with highest accuracy at a 30-sample patch size and 15% masking ratio, and forecasted 15-second FHR horizons from 30-minute contexts.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: