Conformal prediction quantifies wearable cuffless blood pressure with certainty

Though wearable cuffless blood pressure (BP) measurement technology has attracted significant attention from both academia and industry, the ability of existing methods and devices to track dynamic BP changes and provide reliable BP readings remains low, especially in ambulatory environments. This study develops and validates an algorithm for 24-h ambulatory cuffless BP confidence intervals (CIs) estimation with conformal guaranteed coverage of the true BP values using wearable electrocardiogram (ECG) and photoplethysmogram (PPG) on subjects in the ambulatory setting. First, a quantile loss-based Gradient Boosting Regression Tree (GBRT) model was trained to obtain ambulatory BP estimates along with model uncertainty. The model uncertainty was then calibrated using conformal prediction to obtain CIs with guaranteed reference values coverage. Ambulatory physiological data from 483 participants from the Aurora-BP study dataset were used for model validation. For ambulatory measurements during the daytime phase, the mean absolute difference (MAD) of the systolic BP (SBP) and diastolic BP (DBP) estimated by the proposed model was 14.32 mmHg and 9.53 mmHg, respectively. For ambulatory measurements during the nighttime phase, the MAD of SBP and DBP estimated by the proposed model were 14.22 mmHg and 10.13 mmHg, respectively. Providing CIs with guaranteed reference BP coverage for 24-h ambulatory BP estimation can enhance the trust of patients and physicians in wearable devices, thereby facilitating the prevention, screening, and management of hypertension.