An optimized framework for the prediction of blood pressure based on morphological and dynamic features of PPG and ECG

Tai, Chen Boon (2024) An optimized framework for the prediction of blood pressure based on morphological and dynamic features of PPG and ECG. Project Report. Universiti Teknikal Malaysia Melaka, Melaka, Malaysia. (Submitted)

Text (Full Text) An optimized framework for the prediction of blood pressure based on morphological and dynamic features of PPG and ECG.pdf Download (4MB)

Abstract

Continuous blood pressure (BP) monitoring is crucial for managing hypertension. However, current methods have drawbacks such as risks and discomfort for the patient. To address this, some studies have explored BP prediction through PPG and ECG signals. This project aims to develop two BP prediction models: one for systolic and one for diastolic pressure, by identifying and extracting BP-related features from PPG and ECG signals along with demographic features, using a machine learning model and Shapley Additive Explanations (SHAP). The model's performance is evaluated against AAMI and BHS standards. Two experiments were conducted, which includes identifying the best machine learning model and determining the best feature combination for BP prediction. Initially, features were extracted, and both Support Vector Regression (SVR) and Random Forest models were trained on the dataset. The results from model selection show that Random Forest performs better than SVR, hence, it is used to develop the BP prediction models. The results from feature analysis reveal that both signals and demographic features contribute to BP prediction. The inclusion of ECG signals and demographic features is found reduces the Mean Error (ME) of prediction by approximately 24.13% for SBP and 81.50% for DBP compared to using only PPG signal. In this project, SHAP feature selection is introduced, which involves ranking features according to their importance in machine learning model predictions, followed by an iterative process of removing the least important features to select the optimized feature combination based on the lowest root mean square error. The optimized feature combination is then used to develop the final BP prediction model. The final result indicates that SHAP feature selection managed to reduce the number of features used in SBP and DBP models by up to 48.72% and 50%, respectively, while still providing comparable results to the models with the full set of features. This output is expected to be beneficial for medical teams in clinical studies on blood pressure and cardiovascular diseases.

Actions (login required)

View Item