Blood pressure estimation, from PPG signal, using machine learning algorithms

Abstract

The aim of this thesis work was to implement a methodology for the prediction of systolic and diastolic blood pressure starting from the PPG signal, with the characteristics of non-invasiveness, continuity and remote usability.

As a result of the concatenation of different datasets in terms of activity, PPG signal acquisition mode and blood pressure measurement, the proposed approach presents considerable heterogeneity. As a result of the different approaches for feature selection, the proposed models show that all the domains analysed are important in the pressure estimation.

The analysis of the results of the models on systolic pressure shows at the end of the optimisation of the hyperparameters that the model that shows the best performance is the one based on the Gaussian Process Regression algorithm. This algorithm also appears to be the best for the estimation of diastolic pressure, the results of which are in agreement with all the proposed standards. For these two models, a study was carried out on their interpretability through the analysis of the importance of the features and it emerged that the most important features for both models are those related to Pulse Transit Time, a fundamental feature for the implementation of models based on PPG signal that has a strong clinical relevance. This analysis of the most relevant features and their clinical relevance shows that although the GPR is a poorly interpretable and explicable model, the proposed models exhibit good reliability. The analysis of the borderline cases showed that for both the systolic pressure and diastolic pressure models, the reason for the error is to be found more in the relevant weight of less important features than in the training model and in the low score of some features with greater weight. This highlights both the hierarchy of features in our models but also, albeit with their weight, that all features are relevant in the prediction of pressure values. In conclusion, this thesis work proposes an approach that is new in the literature, very generalisable and valid for estimating systolic and diastolic pressure. The most relevant features are those relating to the non-invasiveness of the methodology and the possibility of remote recording for patients. The proposed models can certainly be optimised; in fact, in the future, tests could be carried out on new datasets with a higher numerosity.

Objective

Development of a methodology, with the characteristics of non invasiveness, continuity and remote usability for monitoring systolic and diastolic pressure from the PPG signal.

Research methodology

• Find public Dataset (PPG signal)
• Processing: Sampling, Filtering (Butterworth, continuous wavelet transform), peak detection, HRV feature extraction
• Dataset Concatenation
• Training of five classification algorithms (Multiple Linear Regression, Random Forest, Support Vector Regression, Least- Square Boosting, Gaussian Process Regression) in Matlab

• Analysis of feature importance

Conclusions

The model that shows the best performance is the one based on the Gaussian Process Regression algorithm. This algorithm appears to be the best for the estimation of diastolic and systolic pressure, the results of which are in agreement with all the proposed standards. For both regression models, it emerged that the most important features are those related to Pulse Transit Time, a fundamental feature for the implementation of models based on PPG signal that has a strong clinical relevance.

Future developments

The proposed method allows the use of a non-invasive methodology and gives the possibility of remote registration for patients. The models used can certainly be optimised by increasing the numerisity of the dataset to improve its accuracy.