Blind Source Separation for Fetal PPG with Rate-Based Proxy Supervision

Separating fetal photoplethysmography (PPG) from non-invasively acquired transabdominal recordings is a critical step toward reliable estimation of fetal arterial oxygen saturation (fSpO$_2$). We present an end-to-end attention-based framework that extracts fetal components from mixed maternal and fetal signals. Since ground-truth fetal PPG cannot be collected in practice, the model is pre-trained on physics-informed synthetic mixtures and evaluated on in-vivo data from pregnant ovine studies. To enable deployment in real-world settings, we employ a multi-objective design in which separation is paired with rate estimation, serving as a proxy for assessing separation quality when ground truth is unavailable. On in-vivo recordings, the proposed framework improves downstream fSpO$_2$ estimation, reducing mean absolute error by 33.3\%, lowering the standard deviation of error by 30.7\%, and increasing correlation with reference values by 26.9\% compared to baseline methods. Overall, this end-to-end pipeline has the potential to enhance real-time fSpO$_2$ estimation and advance non-invasive fetal monitoring in clinical practice.