| dc.description.abstract | According to the 2022 analysis of the top ten causes of death in Taiwan, the total number of deaths caused by cardiovascular-related diseases has surpassed those caused by the most lethal cancers, including lung and liver cancer. Hypertension, in particular, is a leading cause of cardiovascular diseases and mortality in developed countries. Continuous blood pressure monitoring is crucial for the effective prevention of cardiovascular diseases.
This study focuses on predicting continuous diastolic and systolic blood pressure using the photoplethysmography (PPG) signal generated by a single heartbeat. It explores the results of four machine learning models for blood pressure prediction and compares three common continuous blood pressure detection methods: PPG+Electrocardiography (ECG), PPG alone, and pulse transit time (PTT). The study further analyzes the significance of 32 selected features for blood pressure prediction and compares the most suitable features and methods for predicting diastolic and systolic blood pressure.
Our research found that the linear regression and random forest regression models are highly flexible, allowing for comprehensive prediction of more dispersed blood pressure distributions. Among the feature groups, PTT is the only method that does not use feature engineering, resulting in lower accuracy compared to PPG+ECG and PPG methods. The best results for diastolic pressure prediction were obtained using the PPG method, with a mean absolute error (MAE) of 2.21 mmHg and a standard deviation of 3.6, after removing the 25% least significant features. The best method for systolic pressure prediction was PPG+ECG without feature removal, achieving an MAE of 4.2 mmHg and a standard deviation of 6.96. Additionally, we found that the width of the ascending height is suitable for systolic pressure prediction, the width of the descending height is suitable for diastolic pressure prediction, and systolic pressure prediction requires ECG-related features more than diastolic pressure prediction. | en_US |