Application of Deep Neural Network Models for Blood Pressure Classification based on Photoplethysmograpic Recordings

The measurement of blood pressure (BP) in an uninterrupted and comfortable way for the subject is essential for early diagnosis and monitoring of cardiovascular diseases (CVD). In fact, hypertension is the main risk factor for CVD because, being a hidden health problem with no symptoms until late stages of the disease are reached. This work investigates whether deep neural network models are able to discriminate between healthy and hypertensive subjects based on photoplethysmographic (PPG) recordings, without the need of electrocardiographic (ECG) recordings as well as avoiding manual morphological feature extraction, as has been popularly used in many previous studies. Recordings analyzed consisted of 635 simultaneous PPG and arterial blood pressure (ABP) signals from 50 different patients. The classification was performed with GoogLeNet, ResNet-18 and ResNet-50 pretrained convolutional neural networks (CNN) using as input images the scalogram of PPG segments obtained by continuous wavelet transformation (CWT). Additionally, Adam and SGDM training solvers were used to compare classification performance. After applying early stopping to avoid overfitting, training was performed with more than half of the epochs using Adam optimizer. ResNet-18 CNN provided the highest classification performance with sensitivity of 95.68%, specificity of 93.65%, F1-score of 95.61% an Area under the Roc area of 98.77%. Hence, the application of deep neural network classification models using time frequency transformation of PPG recordings has been able to provide outstanding results in blood pressure classification without requiring neither morphological feature extraction nor ECG features.