Heart Sound Analysis for Discrimination of VSD

A ventricular septal defect (VSD) is the most common congenital heart disease, which can be cured with a high probability if it is detected in an early stage. In our previous researches on heart sounds (HSs) analysis, the detection methods of heart disease using the cardiac sound characteristic waveforms in time domain or in frequency domain were proposed, and have been succeed in discriminating several heart murmurs. In this paper, we are going to apply these methods to detect VSD. Based on analysis results, a new approach by using the feature parameters both in time domain and in frequency domain is proposed to achieve higher discrimination rates.


Introduction
A VSD is one of the most common congenital heart disease, accounting for 0.24% of newborn babies [1,2], and it can be cured with a high probability if this disease is detected in an early stage. Recently, with the high development of computer technique and digital signal processing technology [3,6], more and more researches are concerning on the HSs analysis. Through comparative analysis, the simple and effective methods of our previous studies [4,6] were proposed to detect heart murmurs with higher classification accuracy. In this study, we are going to use our previous method to detect VSD from normal cases. According to experimental results, a new approach is proposed to achieve higher discrimination rates. The rest of the paper is organized as follows. Section2 introduces HSs acquisition and preprocessing. Section3 proposes HSs analysis method. Section4 presents discrimination analysis. Finally, the conclusions are summarized in Section5.

HSs Acquisition and preprocessing
Auscultation denotes the act of analyzing sounds in the body that is produced in response to mechanical vibrations generated in the organs. Therefore, for different heart murmurs, we should analyze HSs collected from different auscultation areas, while for VSD cases, it is reported that the HSs collected from tricuspid area can supply more important information [8]. In this study, analyzed HSs were collected from tricuspid area by the HSs acquisition system, meanwhile, sampling frequency Fs were set as 44.1 kHz. The basic HSs consist of two primary components which are often described as the first heart sound (S1) and second heart sound (S2). To analyze the useful information, wavelet decomposition (WD) is used as pre-processing for cancellation of the unwanted frequency components over 700Hz and below 20Hz. Daubechies type wavelet DB10 is used as a mother wavelet. Finally, the filtered signals, x(t) with 21.5-689Hz is gained.

Cardiac sound analysis
In this section, firstly, the diagnostic features, [T11,T12] from characteristic waveform (W t ) in time domain as detecting character of HSs are extracted. And then to experimental analysis, the diagnostic features [Fg,Fw] extracted from the frequency characteristic waveform (W f ) are proposed to detect VSD. By analyzing the distribution of [T11,T12] and [Fg,Fw] for VSD and normal HSs. Finally, a new approach is proposed for detecting VSD and normal cases.

Time domain analysis
In our previous studies, time characteristic waveform analysis method [4] could realize detection of several heart diseases. So this method is used to analyze performance evaluation for discriminating VSD HSs and normal HSs, and the detailed is in following.

Characteristic waveform(W t ) extraction
Consider a data series x(t), t=1,2,…,N, by WD for HSs, where N denotes the number of data. Then the W t based on the Viola integral method is given by . ( At last, the normalization is applied by setting the amplitude of W t within 1.0. Since many experiments show that the duration of S1 or S2 is over than 0.06 second [4,8], we set δ=0.03×Fs=1323. As an example, Fig.1 plots x(t) daubed with gray and it's W t . Fig.1(a) shows the case of normal sound and Fig.1(b) is VSD sound. As for the normal sound, W t seems to have longer time interval between two abutted S1 and between S1 and S2 than VSD.

Diagnostic features definition and representation
As mentioned above, a concept for defining the diagnostic parameters is described in Fig.1(a) and (b). A threshold value (Thv1) is selected first at a suitable value, the time intersection between the crossed points of the W t on the Thv1 line are defined by a(S1 i ),b(S1 i ), a(S2 i ) and b(S2 i ) (i=1,2,.,N) in a sequential order as shown in Fig.1. The center of gravity, especially denoted G t (S1 i ) and G t (S2 i ) as shown in Fig.1(a) and (b), the time index of G t (S1 i ) and G t (S2 i ) are gained by T11 i is the time interval between G t (S1 i ) and G t (S1 i+1 ),T12 i is the time interval between G t (S1 i ) and G t (S2 i ).To make the parameters [T11 i , T12 i ] visually, a two-dimensional plot, scatter gram, on [T11 i , T12 i ] is introduced as shown in Fig.1(c).

Experimental results and discussions
To validate the proposed method, in this paper, the used data set with 468 sound samples that consisted of 242 normal and 226 VSD cases. The normal cases were from 23 health students in university, the VSD cases were from 17 patients in hospital.

Case of normal HS
According to the distribution of [T11,T12], generally, normal HSs could be divided into three types. Fig.2(a)

Case of VSD HS
However, for VSD cases, according to the strength of noise and the heart beat, VSD case generally can be divided into three types. The first type is that S1 and S2 can not be distinguished due to stronger noise which causes [T11,T12] are far bigger than NM case, just as a VSD1 case collected from a male with VSD of age 12 and 32kg (Fig.2(d)). The second is that the noise almost not affect the [T11,T12], just as VSD2 from female of age 7 with weight 30 kg (Fig.2(e)). The third is the

Summary
Analysis results show that there is the common region between VSD cases and NM cases, which includes 32.6% of NM cases and 29.1% of VSD cases, so it's impossible to distinguish VSD cases from NM cases. Next, frequency analysis method is proposed to detect normal and VSD cases.

frequency analysis method
The envelope curve method in frequency domain has been proved to recognize several heart murmurs in our previous studies [6].Based on this point, a new characteristic waveform curve in frequency domain (W f ) is proposed to detect VSD. Consider a data series x(t),t=1,2,…,N, where, N denotes the number of data. Then W f (f=1,2,..N) based on the Viola integral method is given by . ( By many experiments, δ is set as 8. At last, the normalization is applied by setting the amplitude of W f within 1.0. As an example, frequency distribution daubed with gray and W f are plotted in Fig.3. Fig.3(a) shows a normal case and Fig.3(b) is a VSD case. W f of the normal sounds, which has the lower density frequency component focused on a narrower region compared with VSD cases.

Features extraction and representation
Two diagnostic features, Fg and Fw are defined, which correspond to the frequency index of the center of W f during interval [0,(N-1)/2] and the frequency width of W f on a Thv2 as shown in Fig.3(a) and (b).The representation of [Fg,Fw] is introduced in Fig.3(c). Here, Fg is gained by . By many experimental analysis, when Thv2 is selected in the interval [0.1,0.2]. Generally, there are greater differences between NM cases and VSD cases. In this study, the Thv2 is set at 0.2.

Experimental results and discussions
Character of W f as shown in Fig.4(a-f). Fig.4(g)