FHRFormer: A Self-Supervised Masked Transformer Framework for Fetal Heart Rate Time-Series Inpainting and Forecasting

Researchers propose FHRFormer, a masked transformer-based autoencoder that reconstructs missing fetal heart rate data from wearable monitors using self-supervised learning. The method addresses signal dropout caused by sensor displacement and positional changes, preserving spectral characteristics better than traditional interpolation while enabling both data inpainting and forecasting for improved fetal risk assessment.

FHRFormer represents a significant advancement in medical signal processing, tackling a persistent challenge in prenatal monitoring where approximately 5-10% of newborns require breathing assistance at birth. The core innovation lies in applying masked transformer architectures—a self-supervised learning paradigm proven effective in language and vision domains—to healthcare time-series data. This cross-domain application demonstrates how modern deep learning techniques can address domain-specific problems in medical settings.

The research emerges from a practical need in obstetric care. Current wearable fetal heart rate monitors generate continuous data streams, but maternal movement and positional changes create signal gaps that undermine the reliability of AI-based risk prediction algorithms. Traditional interpolation methods fail to preserve the temporal and spectral characteristics essential for accurate clinical analysis. FHRFormer's masked autoencoder approach learns to reconstruct missing segments by understanding underlying patterns in complete segments, effectively learning the signal's natural structure.

The framework's applicability extends beyond retrospective analysis of research datasets. Its potential integration into wearable devices could enable real-time signal recovery and risk detection during labor, supporting earlier obstetric interventions when outcomes matter most. The robustness across varying missing data durations suggests practical deployment flexibility across different monitoring scenarios.

Looking forward, the success of this approach validates transformer-based architectures for medical time-series problems where data scarcity and signal quality represent persistent challenges. Future developments might include integration with clinical decision support systems and validation across diverse patient populations to establish diagnostic reliability standards.

• →Masked transformer autoencoders effectively reconstruct missing fetal heart rate signals while preserving spectral characteristics superior to traditional interpolation
• →The method handles variable-duration data gaps, enabling both signal inpainting and forecasting for robust clinical analysis
• →Real-time integration into wearable FHR monitors could improve early detection of fetal distress requiring breathing assistance
• →Self-supervised learning approach reduces dependency on labeled medical data, addressing a critical constraint in healthcare AI development
• →Framework validation on research datasets supports development of AI-based obstetric risk algorithms with improved data quality