A Wavelet-based Cuffless Blood Pressure Algorithm To Detect Acute Hypertension With Wearable Devices

Acute hypertensive (HT) emergencies such as malignant hypertension and autonomic dysreflexia must be treated rapidly to prevent end organ damage. Therefore, wearable devices which rapidly detect HT emergencies draw significant interest. However, there are no clinically approved methods to detect blood pressure from wearable devices. PURPOSE: Therefore, the purpose of this study was to develop an algorithm for wearable devices that can rapidly detect HT events in real time. Our hypothesis was that our wavelet-based algorithm would detect acute HT events in young-healthy adults. METHODS: 43 young-healthy (24 ± 4 yrs) subjects participated in this study. Electrocardiogram (ECG) and continuous blood pressure (BP) using finger photoplethysmography (PPG) were collected during 10-min of supine rest and 1-min of handgrip exercise at 50% maximum voluntary contraction to elevate blood pressure. The novel algorithm was used to calculate pulse arrival time (PAT) from ECG and the PPG sensors. PATs were converted to z-scores to be comparable between subjects. HT events were defined as BPs greater than three standard deviations from the mean resting BP. Subjects were subdivided into training (n = 24) and testing (n = 19) sets. A logistic regression model was fitted to the training set and its performance was assessed on the training and testing sets. Model performance was assessed with precision-recall area under the curve (PR-AUC) and F1 scores (the harmonic mean of precision and recall). 95% confidence intervals (CIs) were generated for PR-AUCs and F1 scores by bootstrapping. PR-AUCs and F1 scores for a random classifier were also determined. Data are presented as means and 95% CIs. RESULTS: The PR-AUCs were 0.67 (0.55-0.76) and 0.67 (0.56-0.80) for the training and testing sets, respectively. The PR-AUC for a random classifier was 0.10 (0.09-0.11). The maximum F1 scores were 0.61 (0.52-0.70) and 0.67 (0.58-0.77) for the training and testing sets, respectively. The F1 score for a random classifier was 0.18 (0.16-0.20). CONCLUSIONS: Our novel algorithm can be used to detect NT and HT events, and our machine learning model may be well-generalizable. Overall, our novel algorithm may be useful for monitoring HT emergencies with wearable devices. Supported by: This research was supported by the National Institutes of Health (R01HD106911, R01AG0778031, P20GM152301) and the Nebraska Research Initiative.