Recurrent Neural Network Models for Blood Pressure Monitoring Using PPG Morphological Features

Continuous non-invasive Blood Pressure (BP) monitoring is vital for the early detection and control of hypertension. However, this is yet not possible as all current non-invasive BP devices are cuff-based devices and hence precluding continuous monitoring. Several methods have been proposed to overcome this challenge, one of which utilises the Photoplethysmograph (PPG) signal in an effort to predict reliable BP values from this signal using various computational approaches. Although, good performance has been reported in the literature, it was mainly achieved on a small inadequate sample size using conventional models that are unable to account for the temporal variations in the input vector. To address these limitations, this paper proposes cuff-less and continuous blood pressure estimation using Long Short-term Memory (LSTM) and Gated Recurrent Units (GRU). The models were evaluated on 942 patients acquired from the Multiparameter Intelligent Monitoring in Intensive Care (MIMIC II) dataset. The proposed models produced superior results in comparison with conventional artificial neural network. In particular, the best performance was achieved by the GRU, with mean absolute error and standard deviation of 5.77 ± 8.52 mmHg and 3.33±5.02 mmHg for systolic (SBP) and diastolic blood pressure (DBP), respectively. Furthermore, the results comply with the international standards for cuff-less blood pressure estimation.