A machine learning approach for non-invasive blood pressure estimation

The increase in the production and adoption of wearable devices in everyday life, as well as the improvement of their refinement and precision, encourage the development prospects of this market in the future. The wearable market size in 2022 is estimated to reach 57,653 million dollars. Due to its continuous increase in easy to use, flexibility and convenience, it can grant the possibility of returning the collected data in real-time. Nevertheless, there are currently limits due to the life of their battery which is holding back their position on the market. In parallel, the increase in their computational capacity and the increase in the simplicity with which it is possible to integrate them with complex algorithms, such as machine learning algorithms encourage the study of new methods for estimating vital parameters with non-invasive techniques, in the topic of this work the blood pressure measurement. For decades, cardiovascular diseases (CVD) have been a major cause of mortality and morbidity around the world [Can Photoplethysmography Replace Arterial Blood Pressure in the Assessment of Blood Pressure?] and the number of patients suffering from CVD increases year by year. One of the major risk factors for the development of CVD is chronic hypertension, which is characterized by elevated baseline blood pressure over long periods of time. Therefore, estimating blood pressure assumes a fundamental role in monitoring the health status of the population. Currently the measurement takes place mainly through auscultatory devices such as the sphygmomanometer, which is digital or classic, requiring in the second case the presence of a trained operator, and they require static conditions for the subjects. Encouraged by this, with the increase in the refinement of wearable devices and the implementation of machine learning algorithms, it has been explored other physiological parameters that the possibility to measure blood pressure via indirect measurement. An example is the Pulse Wave Velocity (PWV) study, in which the variation in the velocity of the pressure wave produced by the passage of blood in the vessels is observed. It is associated with other observations related to the quantity of blood and the variation in diameter of the vessels and can be converted to the pressure measurement that corresponds to a specific observation window (in our case 10 seconds). In this work, a model was implemented to estimate the pressure through the collection of data from online databases (MIMIC data []) for the identification of the best characteristics of the signals taken and the training of a linear regression algorithm. Those data estimate the pressure from electrocardiogram (ECG) and photoplethysmography (PPG) signals to identify the best morphological or temporal characteristics for training a linear regression algorithm through which to estimate blood pressure values. The results achieved by implementing the Random Forest as a linear regression model and splitting the dataset into training and test set showed a quadratic error of approximately and standard deviation of 7 ± 9 mmHg for SBP, 5 ± 7 mmHg for DBP and 5.5 ±7.5 mmHg for MAP. Finally, the system was validated through 10fold cross-validation.