A computationally efficient CNN-LSTM neural network for estimation of blood pressure from features of electrocardiogram and photoplethysmogram waveforms

Abstract

Continuous blood pressure (BP) monitoring would significantly improve diagnosis and treatment of hypertension. Current at-home monitoring relies on uncomfortable and unreliable cuff-based devices, which are incapable of continuous measurement. In this work, we present a new hybrid neural network (NN) that combines convolutional layers with long short-term memory (LSTM) layers to classify systolic blood pressure (SBP), diastolic blood pressure (DBP) and mean arterial pressure (MAP), using 12 straightforward features extracted from electrocardiogram (ECG) and photoplethysmogram (PPG) waveforms. Our proposed network achieves mean absolute errors (MAEs) of 4.53 mmHg, 3.37 mmHg and 3.36 mmHg for SBP, DBP and MAP respectively. Additionally, our scheme passes the criteria outlined by the Association for the Advancement of Medical Instrumentation (AAMI) and achieves an A grade in accordance with the British Hypertension Society (BHS) protocol. These results provide a deep learning approach to BP estimation that could be implemented in low-power wearable devices.