Advancing Cuffless Arterial Blood Pressure Waveform Estimation: Time-Series Deep Neural Network Approach

Existing deep learning models for arterial blood pressure (ABP) estimation are becoming increasingly complex. They mainly treat the estimation as a sequence-to-sequence (seq2seq) task, to establish the relationship between input physiological signals and the corresponding BP within the same time frame. However, this approach may overlook the rich temporal information embedded in physiological signals. In this study, we propose a time-series training strategy for ABP waveform prediction. We compared two deep learning models of different sizes - the smaller gMLP and the larger UtransBPNet - in both seq2seq and time-series training ways. The findings indicate that, the models trained with the time-series method achieved significant enhancements in performance compared to their seq2seq counterparts, with mean absolute error (MAE) reductions of 2.0 and 0.9 mmHg for gMLP and UtransBPNet, respectively. This improvement was more pronounced in the smaller, simpler-structured gMLP network. Additionally, the time-series training approach exhibited superior predictive abilities for out-of-distribution data. In conclusion, this straightforward time-series approach offers a novel perspective for developing efficient models for cuffless arterial BP estimation, making it a promising candidate for implementation in edge wearable devices.