Blood pressure monitoring with piezoelectric bed sensor systems

Abstract

ObjectiveBallistocardiography (BCG) measures vital signals without direct contact, which shows great potential for continuous sleep monitoring. This study proposed a BCG-based blood pressure (BP) estimation algorithm framework using piezoelectric bed sensor systems. Methods: To derive BP, a combination of morphological features, spectral features, and fractal dimensions of the BCG signal were utilized. Bayesian neural networks were employed to weigh the contribution of each feature and generate input-dependent coefficients. Two data balancing procedures were tested, and the proposed system's effectiveness was evaluated, including in-hospital patients and healthy subjects. Transfer learning techniques were employed to further improve the system's performance and showcase the similarity between bed systems with different piezoelectric sensors. Results: The combination of morphological and spectral features significantly improves BP estimation accuracy. The fractal dimension captures short-term BP fluctuations, improving intra-subject BP trend estimation. For young and healthy subjects, calibration-free mean absolute error (MAE) for systolic BP (SBP) and diastolic BP (DBP) is 4.20/4.25 mmHg at a 5-second time resolution. In the case of in-hospital patients, the best MAE for overnight SBP/DBP is 9.96/7.59 mmHg. Transfer learning, combined with data balancing techniques, substantially enhances BP estimation accuracy for in-hospital patients, providing insights for future algorithm designs. Conclusion: The study concludes that bed-sensor systems can track BP changes in relatively healthy subjects. However, BCG alone may not be sufficient for subjects with severe cardiovascular dysfunctions to obtain reliable BP readings. Transfer learning and proper data balancing may facilitate the fast development of new bed sensor systems with similar technologies.