A Machine Learning Approach to Predict Arterial Blood Pressure from Photoplethysmography Signal

Felipe Meneguitti Dias1, Thiago Costa2, Diego Cardona Cardenas2, Marcelo Toledo1, Jose Eduardo Krieger3, Marco Gutierrez3
1Instituto do Coração, Hospital das Clínicas HCFMUSP, Faculdade de Medicina, Universidade de São Paulo, 2Heart Institute, Clinics Hospital, University of São Paulo Medical School, 3Heart Institute University of Sao Paulo


Blood pressure (BP) monitoring is a basic procedure for the physiological measurement of the cardiovascular system, especially because high BP, although preventable, is a major risk for stroke, heart failure, and other serious conditions. Photoplethysmography (PPG) is a promising technology developed to allow non-invasive, regular, or even continuous measurement of blood volume variation. Recently, some works have tried to use PPG signals to estimate BP. In this work, we propose a regression model based on the Category Boosting algorithm (CatBoost) that uses 133 morphological and temporal features from the PPG signal to estimate the corresponding diastolic and systolic BP. We processed and selected a total of 50,182 windows of 1,000 samples (sampling rate of 125Hz during 8 seconds) of PPG and BP signals from the MIMIC-II dataset, distributed into training and test sets. Three different data cross-validation schemes were adopted. The model prediction metrics were evaluated by Mean Error and standard deviation (ME[STD]), and Pearson's Correlation Coefficient (R-value). For one of the validation schemes, we obtained, for the diastolic BP, 0.02[3.77] mmHg with an R-value of 0.93; and for systolic BP: 0.05[7.84] mmHg with an R-value of 0.93. Our results meet the AAMI standard and are comparable to the state of the art. However, we show that these results rely on a specific validation scheme.