City Research Online

Abstract

Blood pressure (BP) is one of the four primary vital signs that provides important information regarding patients' cardiovascular system conditions. Continuous and regular blood pressure monitoring is essential for the early diagnosis, prevention and management of cardiovascular disease (CVD) and haemodynamic diseases (hypertension and hypotension). Current clinical blood pressure measurement techniques are either invasive or cuff-based, which can be impractical, intermittent, and uncomfortable for patients during frequent measurements. Considering these challenges, several studies have suggested new non-invasive and cuffless blood pressuring measurement techniques using physiological signals, such as, the Electrocardiogram (ECG) and the Photoplethysmogram (PPG). In particular, indirect cuffless BP measurement techniques using pulse transit time and pulse arrival time have been extensively investigated over the last few decades. However, these techniques require two measurement sensors, frequent calibration, and hence, they are also impractical and inconvenient for continuous BP measurements. More recently, with the advancement of computational techniques, including machine learning and artificial intelligence, a new simple and innovative approach using only PPG signals have been proposed in the literature for cuffless and continuous monitoring of blood pressure. However, the majority of these studies have been unable to achieve acceptable accuracies that comply or satisfy the international standards for cuffless BP monitoring. Thus, further investigations are required to realise this approach.

In this research, a total of 52 features have been extracted from the PPG and their individual impact on BP have been rigorously evaluated using several statistical and machine learning techniques. As a result, only the most important features for estimation of BP were selected, effectively reducing the input vector by more than half. Two datasets were created to accommodate the two input feature vectors. The PPG and reference BP signals were derived from the publicly available MIMIC II database. In order to estimate BP, a total of nine machine learning and neural network models have been implemented and evaluated on the two datasets. Out of the nine models, four are widely used classical machine learning models, and five neural network models, three of which are conventional models and two advanced models have been proposed for BP estimation using only one PPG signal. The results of all these models have also been compared against well established studies in the
literature.

The results obtained using the classical machine learning models, namely, multilinear regression, random forest, adaboost and support vector machine, were poor and inferior to all the neural network models. A slight performance improvement was achieved using the non-recurrent multi-layer perceptron, however, the error was still much higher than the internationally acceptable range. On the other hand, a significant improvement was achieved for the first time by using the recurrent neural network models, namely, Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU). The two proposed neural network models further enhanced the BP estimation accuracies and were able to reduce the mean absolute error (MAE) to a range below 5 mmHg. In particular, the best performing model was the one bidirectional GRU layer, followed by two unidirectional GRU layers, and an attention layer. The obtained MAE and standard deviation (SD) were 4.79+/-8.08 mmHg for systolic BP (SBP) and 2.77+/-4.72 mmHg for diastolic BP (DBP). Furthermore, the DBP estimation were well below the internationally acceptable limits (referring to the AAMI standards of mean error (ME), ME+/-SD less than 5+/-8), while the ME for the SBP estimation were acceptable but the SD exceeded the limits by only 1.34 mmHg.

This research has successfully demonstrated that advanced neural network models
can be used for the non-invasive and cuffless prediction of BP utilising the PPG.

Publication Type:	Thesis (Doctoral)
Subjects:	R Medicine > R Medicine (General) T Technology > TA Engineering (General). Civil engineering (General)
Departments:	Doctoral Theses School of Science & Technology

Export

Downloads