Next-Generation Fetal Heart Monitoring: Leveraging Neural Sequential Modeling for Ultrasound Analysis

Fetal heart rate (FHR) and its variability are crucial indicators of fetal well-being. One-dimensional Doppler ultrasound (DUS) has become a widely used tool for this monitoring purpose, particularly in low-resource settings, due to its affordability, portability, and simplicity. Yet, its potential remains underexplored, with existing methods relying on rigid, non-adaptive algorithms that struggle to capture beat-to-beat variations. This study aims to bridge the gap by delivering reliable estimates through sequential modeling of regions of interest.

We introduce AutoFHR, a novel interpretable neural temporal model based on dilated causal convolutions and attention mechanisms, designed to automatically estimate heartbeat locations within DUS signals. AutoFHR utilizes an innovative learning objective that minimizes generation error while uniquely incorporating a spectral fidelity term to retain the natural rhythm of fetal cardiac activity.

Cross-population, subject-independent evaluations demonstrate AutoFHR's proficiency in heartbeat localization, significantly outperforming conventional methods in FHR estimation while improving FHR variability analysis. AutoFHR achieves a root mean square error (RMSE) of 2.2 beats per minute (bpm) and 2.8 bpm, a maximum limit of agreement of 4.5 and 5.6 bpm, and an estimated bias of 0.3 and 0.1 bpm on the development and external validation datasets, respectively.

Our findings indicate a strong correspondence between estimated and reference fetal electrocardiogram-derived heartbeats and highlight the model's generalizability and robustness over time.

This work advances the clinical utility of DUS-based fetal monitoring by improving FHR analysis, supporting earlier detection of distress and expanding access to quality prenatal care in clinical and remote settings.