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


Abstract

Blood pressure (BP) monitoring is a basic procedure for physiological measurement of the cardiovascular system, especially because high BP, although preventable, is a major risk for stroke, heart failure, myocardial infarction, and other serious conditions. Although lifestyle changes can reduce high BP, about half of individuals are not even aware that their BP is persistently high. Clinical guidelines advise that the sphygmomanometer should be the standard device for measuring BP and that only trained healthcare professionals should perform this assessment. But these requirements have important drawbacks: cuff-based devices make continuous monitoring unfeasible, can cause discomfort or pain, and ambulatory measurements can be influenced by the white coat effect. The photoplethysmography (PPG) has been a promising technology developed to allow non-invasive, continuous and regular measurement of BP, enabling monitoring in the early stages of high BP condition. In this work, we propose a regression model based on the Category Boosting algorithm (CatBoost) to correlate 133 morphological features from the PPG signal with the corresponding diastolic and systolic BP, acquired simultaneously. 87,273 time series (8s window) of PPG and BP signals from MIMIC-II dataset were randomly distributed into training and test sets. A 10-fold cross validation was adopted and CatBoost applied for regression. The model prediction metrics were evaluated by Mean Error Difference (MED) and Pearson’s Correlation Coefficient (R-value), expressed as mean[standard deviation], which resulted for diastolic BP: MED 0.022[4.006] mmHg with R-value 0.891[0.004]; and for systolic BP: MED 0.012[7.869] mmHg with R-value 0.917[0.002]. According to ANSI/AAMI/ISO 81060-2:2013 Failure Criterion 1, the MED mean and standard deviation must be less than 5 mmHg and 8 mmHg, respectively, for both diastolic and systolic BP. Our results meet this criterion, are comparable to the state of the art found in the literature, and indicate that PPG-based BP methods are indeed promising.