1558-1748 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information. This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2019.2962364, IEEE Sensors Journal Identification of S 1 and S 2 Heart Sounds using Spectral and Convex Hull Features Madhusudhan Mishra, Member IEEE, Sawon Pratiher, Student Member, IEEE Hrishikesh Menon and Anirban Mukherjee, Senior Member, IEEE Abstract—A new set of morphological characteristics of Phono- cardiogram (PCG) signal is presented for recognition of first (S1) and second (S2) heart sounds (HSs). Initially, variational mode decomposition on PCG signal generates a set of amplitude and frequency modulated narrow band-limited components (NBCs) and Hilbert transformation of these NBCs comprehends its complex plane analytic signal representation (ASR). Instanta- neous spectral attributes encompassing amplitude modulation bandwidths and convex hull area measure from the ASRs are concatenated to form the feature set. Experimental results on both publicly available and experimentally recorded HSs signals outperform the existing state-of-the-art. Also, the proposed tech- nique does not require any timing information between S1 and S2 and electrocardiogram (ECG) signal reference and is highly robust to noisy real-world PCGs as shown by noise analysis. Index Terms—Auscultation; Convex hull; Heart sounds; Hilbert transform; Variational mode decomposition. I. INTRODUCTION Recent statistics on cardiovascular diseases (CVDs) related deaths shows a rising trend of 31%. Being a pre-dominant source of mortality in today’s world [1], CVDs pose a serious threat to global public health complications due to shortage of medical experts, insufficient infrastructure and poor access to diagnostic services which hinders timely prognosis to patients [2]. Heart sound (HS) carries vital information regarding the state of the heart and manifests variations with pathological conditions [3]. These alterations in phonocardiogram (PCG) signals happen before their symptoms, and hence provides scope for early stage diagnosis [4] [5]. The mechanical vi- brations of the heart generates HSs, as observed during the auscultation process. In healthy grown-up people, two primary HSs appear in a heart cycle which follow certain patterns in terms of the pitch and time of occurrences. The first HS, S 1 (lub) develops at the beginning of the systolic period, because of the closing of mitral and tricuspid valves. The second HS, S 2 (dub) appears at the starting phase of the diastole when pulmonary and aortic valves get closed [6]. The first fundamental HS, S 1 corresponds to a low-pitch signal having longer duration, whereas the second fundamental HS, S 2 is M. Mishra is with the Department of Electronics and Communication Engineering, North Eastern Regional Institute of Science and Technology, Itanagar 791109, India and is pursuing his PhD research in the Department of Electrical Engineering, IIT Kharagpur, Kharagpur 721302, India. (e-mail: ecmadhusudhan@gmail.com). S. Pratiher, H. Menon, and A. Mukherjee are with the Department of Electrical Engineering, Indian Institute of Technology Kharagpur, 721302, WB, India (email: sawon@iitkgp.ac.in, hrishikeshmenon96@gmail.com, and manirban@ieee.org). identified by a comparatively higher pitch and existing for a smaller time-period. In normal cases, the time interval between S 1 -S 2 is shorter than the S 2 -S 1 interval. The presence of third heart sounds, S 3 in adults, and other sounds like heart murmurs are the sign of cardiac abnormalities [7] [8]. Early cardiac abnormalities detection is important for timely intervention and proper treatment of CVDs [2] and the conventional method of cardiac auscultation, i.e., listening to heart sounds using a stethoscope is simple, noninvasive and remains a popular approach for CVDs diagnosis [9] and has shown reliable results in early stage detection of fatal disorders such as acute valvular dysfunction. However, it is subjective, expert- dependent, relying on the skills and experience of the medical professionals and suffers from low diagnostic sensitivity and accuracy in noisy environments Further, high non-linearity and non-stationary nature of HSs with low amplitude and lower frequency components make auscultation a challenging task for medical professionals for precise diagnosis [2]. Moreover, disease like arrhythmia demonstrates special occurrence pat- terns, where normal timing information of HSs components like S 1 and S 2 gets altered. As such primary diagnosis demands a robust automated system. S 1 and S 2 recognition is a crucial and foremost step for any automated cardiac diagnosis system. The reliable automated recognition can effectively help medical profes- sionals in the detection of cardiac abnormalities [10]. Prior art HSs segmentation and detection may be broadly put into two groups: one which uses only PCG signals, commonly known as electrocardiogram (ECG) independent approach, and another is an ECG-dependent approach, requiring ECG as a reference signal. The widely used former approach eliminates recording and analyzing ECG and PCG signals simultaneously and synchronously, which many a times practically difficult. ECG signal-dependent methods extract the information from various components of the ECG signals which include in- stantaneous energy detections, R-wave and T-wave detection [26]. In ECG-independent schemes, various techniques based on Shannon energy [11], wavelet decomposition [12], neural network classifiers [13], HMM [15], moment analysis [17], empirical mode decomposition (EMD) [19], Ensemble EMD [21], and variational mode decomposition (VMD) [25] have been applied for segmentation of S 1 and S 2 . Table I represents summary of selected studies on recognition of S 1 and S 2 components based on the ECG independent approach, the preferred and widely used technique. Most of the available techniques exploit amplitude and timing information between S 1 and S 2 components to achieve good performance. How-