DHF: A Generative AI Approach to Data-Efficient Signal Separation for Wearable Embedded Systems

Separating sensed signals is a crucial task for many wearable systems, where the acquired data contains a mixture of a target signal of interest and several interfering signals, arising from quasi-periodic physiological phenomena such as cardiac and respiratory activity with slowly varying frequency and amplitude. In practice, the task of separating the signals is challenged by quality and quantity of available physiological data. To overcome this limitation, we propose a novel signal separation technique that can effectively isolate the target signal from a sensed mixed signal, using only a single trace of data. We leverage application-specific attributes of the problem, and approach signal separation as an instance of spectrogram in-painting with a custom-designed generative neural network model. The efficacy of the method is demonstrated in the context of transabdominal fetal pulse oximetry application, using both synthesized and in-vivo data collected from pregnant ovine models. Our approach significantly outperforms prior art, achieving 106% average improvement in signal-to-distortion ratio and 85% reduction in mean squared error in synthesized data, compared to the best competing method. In case of in-vivo data collected in animal experiments, where reference signals are unknown, our method substantially improves performance of a downstream task. Specifically, the proposed method reduces the correlation error in estimated fetal blood oxygen saturation by 80.5% compared to a state of the art technique, showcasing its potential to advance wearable systems.