Pedro Pinto da Silva
João Paulo Trigueiros da Silva Cunha
Towards a pervasive system for real-time blood pressure monitoring: Estimation models and events detector
It is widely spread and accepted that regulation and monitoring of arterial blood pressure contributes actively to the prevention of cardiovascular events and diseases.
Although developments have been made in non-invasive ambulatory blood pressure measurement, ambulatory blood pressure monitors or devices based on Finapres technology carry significant costs or performance drawbacks that do not make them eligible for a widespread use of the technology.
Recently, pulse transit time (PTT) based techniques have been suggested as a promising tool to deliver non-invasive and continuous monitoring of BP on a personalized wearable level. In this work the question of how PTT-based models performs as an estimator of SBP is addressed using several approaches. Off-line least squares fitting and recursive least squares are applied to determine parameter correlation and evolution in time. Low correlation coefficients are found and no systematic components are observed. On-line exponentially-weighted regression is applied by performing an initial calibration and then re-calibrating the system every T minutes. Standard deviation of the error and minimum squared error were observed to increase considerably with T, suggesting that initially determined parameters become increasingly obsolete after 30 to 60 minutes of activity.
However, the question whether a PTT-based model is an indicator of SBP variability strong enough to detect events of SBP remained. By imposing medical requirements, namely a threshold for significant SBP changes of 20 mmHg, SBP events are characterized and a set of techniques for SBP event detection are developed. Upon generating the ground truth, the thresholds that maximize the tradeoffs between the number of true and false positive detections are analysed by generating ROC performance curves for three different methods of estimating SBP. Specific points in the ROC curve of a exponentially-weighted regression with period of re-calibration of 20 minutes suggested that the method could be incorporated with current technology for ambulatory BP measurement using a threshold of tr = 12.5 and providing reasonable performance (82% sensibility and 79% specificity).