LOW NOISE HEART SOUND ACQUISITION IN WEARABLE SYSTEM FOR INDIVIDUAL-CENTERED CVD DIAGNOSIS by TAN Zhen Thesis Supervisor: Prof. VAI Mang I Thesis Co-supervisor: Prof. DONG Ming Chui Master of Science in Electrical and Computer Engineering Cardiovascular diseases (CVDs) are the leading cause of global death now and are projected to remain the single leading cause of death in 2030. At present, CVDs are diagnosed by various kinds of hospital-centered tests. Those tests are limited in different aspects. An accessible, cost-effective, convenient and reliable way for long-term and ambulatory early-detection of CVDs by intelligent heart sound (HS) auscultation using wearable system emerges as a promising method for replacing hospital-centered tests. HS acquisition and HS denoising are key for this system, and numerous researchers work on them. However, bottleneck problems exist in those researches. To tackle the bottleneck problems, prototype wearable medical device (WMD) is designed and newly-developed algorithms are proposed. In the thesis, the prototype WMD for HS acquisition including circuit and firmware is designed. The channel to acquire HS and the control and processing unit to handle acquired signal in WMD is the most arduous part and is also the main focus of the thesis. In the design process, rigorous and detail discussion of technique specifications for circuits, chip selection, circuit simulation and test are presented. Practical tests show that the designed prototype operates normally and acquires HS signal right. Yet, a WMD often operates in a noisy and varying environment, thus ambient noise (AN) is an inevitable interference which covers vital HS variants and hinders further diagnosis. Therefore, to acquire low noise HS signal and suppress unpredictable and high-energy ANs is a refractory mission. Various de-noising methods have been researched and developed for HS denoising. However, most methods are not suitable for AN suppression, while interference cancellation (IC) using the least-mean square (LMS) algorithm is a preferable method due to its simplicity and ability of suppressing any stationary AN. Nevertheless the LMS has bottlenecks of low convergence speed (CGS) and low robustness on suppressing abrupt noises. Thus, a novel parallel-training LMS (PTLMS) algorithm is proposed to solve the bottleneck problems. 3600 experiments are performed for both PTLMS and LMS algorithm, and around 80% (2880) of experiments show that the PTLMS algorithm surpasses LMS in terms of CGS and abrupt noise rejection. Furthermore, a few residue noises will still exist after AN suppressing, as well as some body noises and electronic noises. The best scheme for HS denosing is to suppress AN by IC firstly in order to increase HS signal’s SNR, and then suppress residue ANs and other noises by Tang’s method (TM). TM utilizes quasi-cyclic feature of HS and decomposes HS into atoms for fuzzy detection. However, it has bottlenecks of 1) detecting HS periods manually by electrocardiography (ECG), 2) decomposing atoms on time-frequency 2D plane without energy dimension, and 3) deciding density of atoms only by the extent of radius. In the thesis, a proposed ECG-free and automatic HS period detection method called multiply refining algorithm is introduced to settle the first bottleneck problem. By training the algorithm using 11022 labeled peaks first and then testing the algorithm using 7326 labeled peaks, the precision and recall of the test result are near 98% under the tolerated error of 50ms. Cooperating with my colleague Dr. Zeng Ke Han, a cuboid method is developed, which solve the latter two bottleneck problems properly, and this part will not be the main focus of the thesis.