Analysis of pulse wave parameters using supervised machine learning methods
Permanent address of the item is
Pulssiaaltoparametrien analysointi valvottujen koneoppimismenetelmien avulla
Impedance plethysmographic signals recorded with electrodes attached on the human body provide information on hemodynamics and thus the condition of the arteries. The main objective of the thesis is to study if the quantitative analysis of the bioimpedance signals provides additional information on the risk for cardiovascular diseases compared with clinical parameters currently used in the assessment of the cardiovascular risk of an individual. This thesis aims to answer three main research questions: 1) are pulse wave parameters able to evaluate the condition of the arteries, 2) could pulse wave parameters provide information equal to the clinical data, and 3) could the impedance measurements be utilized for cardiovascular risk stratification. This thesis analyzes the bioimpedance signals and clinical data collected in the Cardiovascular Risk in Young Finns Study (YFS). The test subjects were 30–45 years old when the data was collected. In this thesis, both frequency and time domain features including pulse wave decompositions are computed from the pulse waves extracted from the bioimpedance signals and their dependence on clinical phenotypes based on YFS data is evaluated. The YFS data contains demographic information (sex, age), anthropometric data (body mass index (BMI)), clinical information (smoking, hypertension, antihypertensive medication), clinical physiologic data (pulse wave velocity (PWV), blood pressure, heart rate, flow-mediated dilation (FMD)), laboratory analyses (fasting insulin and glucose, lipids of the blood) and imaging data (intima-media thickness (IMT), presence of atherosclerotic plaques in the internal carotid artery). The data was measured from 1853 test subjects, but after removal of test subjects with interrupted measurement or with low signal-to-noise ratio, there are 1738 test subjects used in this thesis. Besides the linear regression analysis, which was implemented to study the association between individual pulse wave signal derived features and clinical reference values, following supervised machine learning methods: linear and quadratic discriminant analysis, support vector machines, naïve Bayes, AdaBoost, Random Forest and k-nearest neighbor are applied to answer the objectives of this thesis. A cross-validation and forward selection are applied to find the most relevant pulse wave features that most accurately classify the test subjects. The results are evaluated with receiver operating characteristics (ROC) curve analysis. This thesis uses three different labeling methods to determine the ground truth for each subject being at low or high risk for cardiovascular diseases: 1) selected cardiovascular risk factors, 2) abnormal body mass index (BMI), blood glucose, triglycerides, high-density lipoprotein (HDL) cholesterol, low-density lipoprotein (LDL) cholesterol or total cholesterol, and 3) existence of atherosclerotic plaque, hypertension or antihypertensive medication. As a result, the calculated pulse wave parameters provide independent information from the clinical data about the condition of the arteries because the combination of pulse wave parameters and clinical data provided the best classification results in most of the cases. However, the calculated pulse wave parameters alone do not provide as good information as the clinical data, which is shown by the fact that the classifying result with only clinical data was better than the classifying result with only pulse wave parameters. As a conclusion, risk stratification improves when the clinical data and the pulse wave parameters are combined. However, the analysis methods of signal processing should be optimized for the bioimpedance measurements. Further in order to verify the classification performance of the developed methods, the data should contain wider spectrum of people, from those who have diagnosed cardiovascular diseases to those who do not have such diseases.