Wrist-worn blood pressure tracking in healthy free-living individuals using neural networks

Blood pressure (BP) is an important indicator of cardiovascular disease. Its measurement is currently done through inflatable cuffs which are obtrusive, especially when used in ambulatory setting and during sleep. This limits the adoption of ambulatory BP measurements in clinical practice, despite that there is strong evidence suggesting that the trends in ambulatory and sleeping BP have unique prognostic value, such as the nocturnal BP dip. In this work an unobtrusive method is proposed to measure BP in ambulatory setting and during sleep. Several physiological characteristics are derived from a PPG sensor worn in a wrist band and combined in a long- and short term memory neural architecture to derive BP trends during day and night. The method is trained on data of healthy free-living individuals with a cuff-based ambulatory BP method as ground truth to predict systolic, diastolic and mean arterial BP. Median Pearson correlation between the proposed method and ground truth was 0.60 for SBP and 0.69 for DBP while mean absolute error was 5.95 mmHg and 4.95 mmHg respectively. The method could reproduce the nocturnal BP dip over a 24-hour period with a Pearson correlation of 0.46. These results show a promising way forward for non-obtrusive ambulatory BP measurement relying on a multitude of physiological signal representations in temporal neural networks.