Confidence-Aware Physiological Parameter Estimation: A Case Study on Cuffless Blood Pressure Monitoring

Abstract

Non-invasive wearable devices introduce more convenient and fashionable alternatives for around-the-clock remote health monitoring at the beat-to-beat resolution throughout the varying contexts of patients’ daily activities - enabling the early diagnosis of life-threatening illnesses. However, as novel wearable modalities emerge, their observed waveforms serve as a proxy to actual hemodynamic behaviors and must be translated to advanced physiological parameters through complex modeling. Data-driven solutions are often preferred since they do not depend on task-specific domain expertise. However, existing approaches pursuing beat-to-beat analysis achieve time-invariant estimation that only considers physiological waveform morphology information from the current cardiac cycle or time window for which a measurement will be assigned. Therefore, these conventional approaches are susceptible to generating inaccurate trends due to data- and model-dependent estimation error for a single or for a series of instances without any ability to self-correct, potentially leading to misdiagnosis by healthcare providers. In this research, using blood pressure (BP) as a case study, we build towards our proposed solution –- Confidence-Aware Particle Filtering (CAPF) -– that for a given instance probabilistically manages multiple hypotheses derived from both time-invariant and also prior information obtained through the construction of trends based on a series of estimated changes. Furthermore, CAPF mitigates erroneous measurements by assigning a confidence score for each hypothesis based on the agreement amongst estimated changes and physiological plausibility. Throughout the included chapters, we first discuss the intermediate steps of automatedly achieving comprehensive feature extraction from the physiological waveforms captured by wearable sensors before exploring the challenges of translating them to BP with time-invariant approaches. Then, we provide the algorithmic details of our proposed CAPF framework and present our experimental results that reflect its superior continuous pulse pressure (PP), diastolic blood pressure (DBP), and systolic blood pressure (SBP) estimation performance compared to conventional approaches. Furthermore, CAPF performs on track to comply with AAMI and BHS standards for achieving a performance classification of Grade A, with mean error accuracies of -0.16 ± 3.75 mmHg for PP (r=0.81), 0.42 ± 4.39 mmHg for DBP (r=0.92), and -0.09 ± 6.51 mmHg for SBP (r=0.92) from more than 3,500 test data points.