Disease-Oriented Hierarchical CVDs Diagnosis Technology through Hemodynamic, Symptomatic and Physiological Parameters by CHEN MuBo Thesis Supervisor: Prof. VAI Mang I Thesis Co-supervisor: Prof. DONG MingChui Master of Science in Electrical and Computer Engineering To cope with the current problems of rapidly growing aging population, pervasiveness of chronic diseases, and soaring cost of personal medical care, a novel, noninvasive, long-term monitoring, and accessible e-Home healthcare is on urgent demand. Specifically, e-Home healthcare is enjoying a rapid expansion to monitor and diagnosis cardiovascular diseases (CVDs) using portable or wearable devices. Comparing with electrocardiogram (ECG) and photoplethysmogram (PPG), sphygmogram (SPG) shows its superiorities of easy to sample, larger signal to noise rate (SNR) and having rich information of vital sign. By using formulas and empirical values based on hemodynamic analysis, a SPG can be extracted to a group of hemodynamic parameters (HDPs). These HDPs have specific biomedical meaning to reflect some aspects of cardiovascular status. Besides, patients’ symptom parameters (SPs) as well as physiological parameters (PPs) are also explored and adopted for CVDs detection. The main ambition of this research is to explore the relationship between HDPs & SPs & PPs so as to detect CVDs for e-Home healthcare usage. As aforementioned, HDPs, SPs & PPs are selected as input features for our cardiovascular system (CVS) due to their effectiveness in clinical practice. 32 HDPs, 5 SPs, and 11 PPs, are used for CVD detection according to the doctor’s suggestion. Nevertheless, we should not cease exploring other features. As an attempt, patients’ birth weight is such a feature to concern. Accordingly, a novel Epigenome-Wide Association Studies (EWAS) is proposed to identify differentially methylated regions (DMRs) relating to birth weight discordance in a whole blood measured sample from 150 pairs of adult identical twins discordant for birth weight. By adapting the co-twin linear mixed model and bumphunter, our proposed method addresses the limitations of existing techniques in EWAS only suiting for case-control samples. Such a combinatory approach, for the first time, is proposed to handle epigenetic data on related twin data and can extend the discordant twin design to more broad and powerful applications. Our analysis successfully detects an interesting genomic region on chromosome 1 differentially methylated for quantitative birth weight discordance, a region harboring two genes (TYW3/CRYZ) which have been reported to associate with insulin resistance, inflammation and risk of type II diabetes and cardiovascular disease. In other words, our finding could suggest that pre-natal condition for birth weight discordance could result in persistent epigenetic modifications potentially implicated in adult health. All in all, birth weight is a powerful indicator for CVD detection but is not adopted in our current e-Home healthcare system due to the restriction of available on-site measured dataset and will be used in our future work. After feature selection, aiming to categorize aforementioned HDPs & SPs & PPs into groups so as to further construct mapping relationship between HDPs & SPs & PPs and CVDs effectively, two disease-oriented HDP&SP&PP categorization methods are proposed, namely, CARTCM and IGAECM. In detail, CARTCM is an efficient intelligent categorization method adapting classification and regression tree (CART). The outcome of CARTCM firms the importance and necessity of disease-oriented categorization; nevertheless, the user manually setting thresholds adopted in CARTCM limits its robustness and stability in practical application. Instead, IGAECM, an intelligent categorization method based on information gain attribute evaluation technique is put forward. Without manually setting thresholds, in IGAECM, an IG-based searching strategy is adopted to group automatically the HDP&SP&PP variables. Hereby, HDPs & SPs & PPs are categorized and the numbers and types of HDPs, SPs and PPs in each of the groups are determined automatically in detecting various CVDs. Such a categorization is able to reduce complexity of inference computation and acquisition cost, increase diagnostic speed, as well as conform to doctor’s clinical diagnosis procedure. To sum up, our proposed categorization methods open a green channel for constructing a multi-label disease-oriented hierarchical CVD diagnosis system. The last of this research is to construct a hierarchical high-efficient acquisition-cost and disease-oriented classifier to diagnose CVDs based on grouped HDPs & SPs & PPs. To tackle such a formidable problem, first, two disease-oriented classifiers, namely, hierarchical probabilistic support vector machine version one (HPSVM.V1) and version two (HPSVM.V2) are addressed here. After that, a rough-accurate hierarchical probabilistic support vector machine (RA-HPSVM) based on aforementioned disease-oriented classifiers is advised. This proposed classifier is assessed and verified by testing on empirical and clinical sampled data sets comparing with state-of-the-art multi-label classifiers. In a conclusion, as a first attempt, such formed classifier not only satisfies acquisition-cost and disease-oriented characteristics of constructing a hierarchical CVD classifier, but also shows good diagnostic performance while reduces the deduction time.