Improvements in Indirect Blood Pressure Estimation via Electrocardiography and Photoplethysmography

Abstract—With the rapid adoption of wearable technology, a ubiquitous platform may exist for continuous cuff-less blood pressure monitoring. Devices that monitor photoplethysmography (PPG) and/or electrocardiography (ECG) signals may be capable of accurately estimating blood pressure. Previous work has demonstrated promising estimation techniques, but they remain infeasible for consumer adoption due to their need for per-subject calibration and/or lack of performance. This paper explores improvements to these approaches by expanding the feature space and exploring more expressive models that estimate mean arterial, systolic, and diastolic blood pressures. These models are trained with features and targets derived from 12,000 sets of ECG, PPG, and blood pressure time-series data. Results indicate that the additional features had a positive impact on results but less so than increasing model complexity. While significant improvements were accomplished, all models struggled to estimate systolic blood pressure to a commercially viable standard. Results indicate that future work should assess convolutional and/ or recurrent neural networks as candidates for this estimation problem.