Innovative cuffless blood pressure estimation using photoplethysmography signal only: development and evaluation of machine learning approaches

Blood pressure (BP) is a vital sign commonly monitored for cardiovascular health. To overcome the limitations of inconvenience and discomfort during frequent cuff inflation, cuffless BP measurement has been investigated mainly based on pulse transit time (PTT), which requires frequent calibration and at least two sensors. BP estimation from a single photoplethysmography (PPG) signal using machine learning algorithms has also been attempted, but the AAMI/ISO standard accuracy has not been achieved. Moreover, these published algorithms have not been evaluated under three different BP categories (Hypotensive, Normotensive, and Hypertensive) and three different body postures. To estimate BP from single PPG signal, three machine learning approaches (Multiple Linear Regression, Support vector machine, Regression Tree) were trained and evaluated on two databases (Queensland and MIMIC-II) which contained PPG and corresponding BPs (SBP and DBP) recordings. Three significant features (total area, rising time, and width 25%) were selected from 17 PPG signal features by multicollinearity test. Two databases were merged into a combined database with less biased BP distribution among three categories, on which three categorically specific BP algorithms were developed from Regression Tree algorithm. Next, a novel two-step approach was developed from K-nearest neighbour (KNN) classification algorithm (to predict BP category) and Regression Tree algorithm (to estimate SBP and DBP). The physiological reliability of the two-step approach was tested on a new database collected from 53 normotensive participants under three different postures (supine, sitting, and stand-up). Regression Tree algorithm achieved the best overall accuracy on the Queensland and MIMIC databases. In terms of the measurement accuracy for each BP category, the acceptable accuracy (referring to AAMI/ISO) was only achieved in Normotensive subjects from the Queensland database. The three specific BP algorithms achieved acceptable accuracy for each BP category of the combined database. The KNN algorithm achieved a 91.7% BP prediction accuracy. The novel two-step algorithm achieved overall (SBP: -0.07±7.1mmHg, DBP: 0.08±6.0mmHg) and categorical accuracy (Hypotensive, Normotensive and Hypertensive, SBP: -3.0±8.2, -0.7±5.8 and 1.7±7.6mmHg, DBP: -0.3±6.3, -1.8±5.1 and 0.5±5.6mmHg, all p>0.05 when compared with reference BPs) on combined data, and on Normotensive data of different postures (supine, sitting and stand-up, SBP (-2.4±7.2,-1.4±6.3 and -0.68±8.4mmHg), DBP (-1.0±6.3, -2.0±6.6 and 0.2±7.9mmHg). A GUI was finally developed for the laptop and a smartphone app that is connected via WiFi to display the estimation results. In summary, this study has developed a novel two-step machine learning approach for BP estimation using the PPG signal only. The algorithm has been comprehensively evaluated and achieved ISO acceptable measurement accuracy.