Although invasive methods are currently used to monitor blood pressure (BP) for intensive care patients, accurate and timely non-invasive BP monitoring in non-invasive way is still significant. Yet, physiological signal data of patients is irregular, with more noise and abnormal patterns included, making accurate and stable prediction challenging. The traditional BP measurement methods are cuff-based, and the prediction accuracy and stability of the machine learning based cuff-less prediction model needs to be further improved. Additionally, data must be cleaned and effective features must be grubbed from the irregular signals, which is a prerequisite for model training.
In the present study, we proposed a novel heterogeneous ensemble learning BP prediction (ELBP) model, where: 1) Related features are systematically extracted and selected for systolic, diastolic and mean BP prediction tasks; 2)Then, multiple regression models are trained and then are weighted for final prediction, wherein the weights are learned from data; 3) Hyper-parameters of each model are optimised using Bayesian optimisation based on cross-validation. We experimentally verified the ELBP effectiveness, the mean absolute error of ELBP is 1.802 mmHg, 3.936 mmHg and 3.121 mmHg for diastolic, systolic and mean BP respectively on mimic-1, and 2.722 mmHg, 5.039 mmHg and 3.812 mmHg respectively on mimic-2. Further experiments demonstrated that ELBP performance is superior to state-of-the-art algorithms on seven evaluation metrics.
In conclusion, BP prediction precision can be further improved by integrating multiple learners appropriately, and this study is valuable in promoting BP prediction in practical application.