Harnessing Photoplethysmography and Deep Learning in Continuous Blood Pressure Monitoring for Early Hypertension Detection

Hypertension (HT) can lead to severe health complications. Therefore, early HT detection is of paramount importance. Photoplethysmography (PPG) stands as a promising solution for blood pressure (BP) monitoring, showing satisfactory but not optimal results. Simultaneously, deep learning (DL) has been hitherto able to improve performance in medical research. The present study aims to harness these powerful tools to assist continuous BP monitoring and detect early signs of HT. 171 five-minute PPG recordings from the MIMIC database were acquired and classified into three categories according to their BP: normotensive (NT), prehypertensive (PHT) and HT. Signals were extracted with a sampling frequency of 125 Hz, resampled to 250 Hz. Each recording was segmented into 2.5-s epochs. Signals were converted to 299 × 299 recurrence plot (RP) images, using m = 3 and τ = 6 ms. Models were trained with the Inception-ResNet-v2 network, using an 80 − 10[%] train-test set. Confusion matrix and performance metrics were extracted in three scenarios: (1) NT/PHT/HT classification, (2) NT/nonNT classification (3) HT/non-HT classification. Classification accuracy, sensitivity and specificity were (1) 96.9%, 91.8% and 99.03%, (2) 97.3%, 95.0% and 98.2% and (3) 75.0%, 72.8% and 76.2%. Classification among the three classes and when focusing on the elevated BP group (non-NT), which is the preliminary HT stage, shows optimal results. HT discrimination is more complicated due to variability in nonlinear dynamics in HT PPG signals. When combined together, DL and PPG can significantly improve performance in elevated BP detection, assisting the inhibition of HT from the first signs of its onset.