Convolutional neural network model for augmentation index prediction based on photoplethysmography

Cardiovascular disease is a prominent cause of death. Among the markers of cardiovascular morbidity, the Augmentation Index (AIx) is the ratio between augmentation pressure and pulse pressure. AIx's increase is associated to vascular stiffness and cardiovascular risk. Currently, AIx is measured employing pressure cuffs reaching the suprasystolic pressure. In order to avoid the use of pressure cuffs and to foster wearable technology capable of assessing vascular diseases, in this study a novel method to predict AIx from multisite photoplethysmography (PPG) through a Deep Convolutional Neural Network (DCNN) model is presented. Seventy-six volunteers (age: 20-80 years) were enrolled in the study. AIx was measured using a commercial instrument (Enverdis Vascular Explorer, VE), whereas PPG was recorded from right tibial, radial and brachial arteries, using a custom-made ECG-PPG system. A leave-one-out cross-validation procedure was performed to test DCNN generalization performances. The DCNN estimated AIx reaching a correlation coefficient between real and predicted AIx of r = 0.74 (p<0.001). Based on the cardiovascular risk provided by VE, a two-class classification (i.e. high- and low-risk) from the cross-validated output of the DCNN was performed. Since the two classes were not balanced, a bootstrap (10000 iterations) was implemented, obtaining an area under the Receiver Operating Curve of 0.93±0.04. Although further studies are necessary to provide a finer classification of the risk (i.e. high-, medium-, low-, very-low-risk) and to exploit the multisite PPG potentialities to early detect cardiovascular pathologies, these results could foster the employment of PPG and DCNN approaches for wearable device-based screenings of cardiovascular risk.