Blood pressure (BP) measurement is an important indication of health and quality of life. However, conventional means of measurement does not provide continual monitoring of data and most of the devices used are cumbersome. Accordingly, developing a non-invasive and cuff-less method for continual BP measurement is essential.
In the development of cuff-less and continual BP measurement, novel methods based on pulse arrival time (PAT) which is obtained from both the photoplethysmogram (PPG) or electrocardiogram (ECG) signals have gained popularity. Although PAT, which is the time delay between the peaks of ECG and PPG signals, is the conventional method generally deployed, still, the results obtained are not consistent and reproducible. Along these lines, the main focus of this research is on the use of physiological signals (which comprises of ECG and PPG signals together with other physiological characteristics) and machine learning methods to predict the diastolic BP (DBP) and systolic BP (SBP) values.
Compared to ECG signals, the collection of PPG signals is simpler and convenient, therefore, novel methods that extract features from the PPG signal are more popular. Nevertheless, previous studies have only focused on the features extracted from the PPG signal and did not consider the physiological characteristics of an individual, which can serve as important predictors for BP. To improve the accuracy in the estimation of BP based on the PPG signal, the first empirical study undertaken in this research extracted features from the PPG signal, and also collated some physiological characteristics (height, weight, and age) of the 191 subjects. The accuracy of BP estimation obtained when prior knowledge of the physiological characteristics are incorporated into the model, are superior to those which do not take the physiological characteristics into consideration. In this empirical study, the best performing algorithm is an artificial neural network (ANN), which obtained a mean absolute error (MAE) and standard deviation (STD) of 4.74 ± 5.55 mmHg for DBP and 9.18 ± 12.57 mmHg for SBP compared to 6.61 ± 8.04 mmHg for DBP and 11.12 ± 14.20 mmHg for SBP without prior knowledge of the physiological characteristics. The best results of DBP estimation complied with Grade A of the British Hypertension Society (BHS) standards, and the implementation of physiological characteristics improved the accuracy of BP estimation.
Some emerging methods have employed complexity features of ECG signals for assessing vital signals, moreover, few studies have also employed the combined complexity features from both PPG and ECG signals for BP estimation. Therefore, the second empirical study conducted in this research was to investigate the performance of a predictive, machine learning BP monitoring system, using complexity features from both PPG and ECG signals. The most accurate DBP result of 5.15 ± 6.46 mmHg is obtained from the ANN model, and the support vector machine (SVM) generated the most accurate prediction for the SBP, which is estimated as 7.33 ± 9.53 mmHg. The best results for DBP fall within the recommended performance of the BHS, but the SBP is outside the range. This demonstrates that the employment of the combination of PPG and ECG signals improved the accuracy of the BP estimation, compared with previously reported results based on the PPG signal only.
Although these findings on the improved accuracy of cuff-less BP estimation show enormous potential, they still cannot meet the official standard. Previous investigations on the adoption of raw signals as input into deep learning for the assessments of vital signals have shown some potential for applications. Few studies have demonstrated the use of raw PPG and ECG signals as input into deep learning models for BP estimation, hence the third empirical study undertaken in this research deployed the use of a novel hybrid model to analyse raw signals for BP prediction. To compare with the methods that were deployed in the aforementioned investigations in this research, traditional machine learning models which utilized morphological and complexity features from PPG and ECG signals respectively, were compared with the novel hybrid deep learning methods. The hybrid model performs best in terms of both DBP and SBP with the results of MAE being 3.23 ± 4.75 mmHg, and 4.43 ± 6.09 mmHg respectively. The estimation accuracy of DBP obtained from this hybrid model is consistent with Grade A of the BHS standard. In addition, all hybrid models achieved lower SBP and DBP errors than traditional machine learning methods. Summarily, the results from the models developed in this research demonstrates practical deep learning models for the prediction of BP which are in agreement with official standards.
|Date of Award||8 Nov 2020|
- Univerisity of Nottingham
|Supervisor||Yaping Zhang (Supervisor), Stephen Morgan (Supervisor), David Cho (Supervisor) & Ricardo Correia (Supervisor)|
- Blood pressure estimations
- machine learning algorithms
- hybrid model