Classification of pulmonary arterial pressure using photoplethysmography and bi-directional LSTM

Abstract

Pulmonary hypertension (PH) is an obstructive progressive disease of lung vascular system which is usually associated with congenital heart disease (CHD) and it induces the right ventricle failure, adversely affecting quality of life and survival. Pulmonary arterial pressure (PAP) is a direct indicator of PH. The existing PAP measurement techniques are invasive and inconvenient, which are not suitable for frequent use or long time monitoring. A noninvasive and convenient method for measuring PAP is essential for the early prevention and diagnosis of PH. The aim of this study was to propose and evaluate a deep learning approach for the classification and evaluation of pH from noninvasive photoplethysmography (PPG) signals. A wavelet scattering transform method was used to automatically extract features from PPG signals. Then a Bi-directional long short-term memory network (Bi-LSTM) was developed to learn extracted features and finally output the predicted PAP classification results. Based on the results from 216 patients’ records, a classification accuracy of 98.52% was presented. This indicated that the Bi-LSTM model trained on features extracted from PPG signals has satisfying performance in PAP classification. This method will enable the noninvasive and painless measurement of PAP. With the development of wearable devices to capture PPG signals from fingertips and the emergence of deep learning models, noninvasive and convenient PAP predictions will greatly contribute to the early diagnosis and prevention of cardiovascular diseases.