1 Introduction

Cardiac diseases are one of the major causes of mortality worldwide. According to World Health Organization (WHO), cardiac diseases are leading cause of death in the world which is about 17.3 million deaths per year [1]. In the developing countries, this situation gets worse due to the lack of medical professionals. However, diagnosis of cardiac diseases at early stages can make the treatment economical, efficient and helpful in preventing death. Heart sound examination is the most common and basic diagnostic technique used by doctors to examine the heart health. Heart sounds are generated by flow of blood through it and beating of heart. Clinically, body auscultations are of crucial importance in determining health of human beings. Auscultation is the act of listening body sounds produced by heart, lungs, blood vessels or other organs using electrical stethoscope [2]. In cardiac auscultation, a doctor may use his stethoscope to hear distinct sounds that provide important data about heart’s condition. A vibration is generated due the pressure of blood flow, opening or closing of heart valve as well as due to the contraction of cardiac muscle. This vibration propagates up through the tissues to the thorax, and it is the method to measure the heart sound. The presence of the problem in heart is identified through murmurs. These murmurs are identified as abnormal heart sound. The turbulent blood flow in the heart system is the reason for murmurs. It is very important to check the pitch and timing of the sounds for diagnosis of the heart condition.

Medically heart auscultation is an essential part for examination of heart to diagnose cardiac diseases at the initial stage. Stethoscope is used for the examination of human heart auscultations. It is an easy, efficient way as well as a computationally cheap technique, but a trained medical professional is needed to understand and interpret heart sounds [3]. Machine learning-based normal and abnormal heart sound classification using audio processing may be a possible method for automatic diagnosis of cardiac diseases or abnormalities without the help of trained medical professionals. Automatic diagnosis of cardiac diseases using heart auscultations maybe of great help in primary health centers for early diagnosis and screening of cardiac disorders.

With reference to audio processing-based diagnosis of heart diseases, some relevant work is summarized in this section. Azmy [4] proposed an algorithm for abnormal and normal classification of heart sounds using new mother wavelet and support vector machines (SVM). The coefficients of new mother wavelet transform are extracted and then these components are ordered using SVM. It is done in two stages, first is discrete wavelet transform which is separated from new mother wavelet transform and then the factual segments are ascertained utilizing SVM. Mayorga et al. [5] had classified heart and lung sounds based on events. Maglogiannisaet.al. [6] proposed an algorithm using support vector machine (SVM) to classify heart valve disease. Gogineni et al. [7] developed a model for heart risk stratification using support vector machine (SVM) and learning vector quantization (LVQ), among which SVM gave 99% accuracy for the prediction of classes with normal, first stroke, second stroke and end of life of patients. Fu et al. [8] had proposed an algorithm for heart sound diagnosis using DTW (dynamic time warping) and MFCC (Mel-frequency cepstral coefficients) on the heart signals. Gudadhe et al. [9] proposed an effective decision support system for cardiac diseases based on MLP neural system technique and SVM. Ghumbre et al. [10] reported an algorithm for heart disease diagnosis on the basis of the support vector machine (SVM) using sequential minimal optimization and for the radial basis function (RBF), orthogonal least square calculation is used. SVM gave the highest accuracy (89.43%) and RBF gave the least (79.30%). Sheela et al. [11] proposed a framework using cardiovascular expectation which is used to predict the sudden cardiac death with the help of the time and frequency of heart rate variability from SVM and ECG. Sonawane et al. [12] had proposed an algorithm for the coronary illness utilizing learning vector quantization (LVQ) neural network calculation. The algorithm is divided into two levels, in the first step, the attributes are treated as the input in the algorithm and then in the next step, LVQ is used to train the algorithms. Pedreira et al. [13] proposed a model on data input selection for the coronary illness.

Table 1 shows a comparative chart of some more existing work in this field of normal and abnormal heart sound classification for diagnosis of cardiac diseases. The dataset used in these various studies are local datasets and not same. The comparison of the different methods should be done on the same database for any scientific merit. Since we have no access to these local datasets, this task remains difficult.

Table 1 Comparative table indicating the performance comparison of existing works with the proposed method
Full size table

Still there is a requirement to design a more accurate and efficient system for automated diagnosis of heart diseases from heart sounds which can work in real time. Automated classification of normal and abnormal heart sound maybe helpful in early diagnosis of cardiac diseases.

The main contribution of this paper is an automatic machine learning-based system for classification of the phonocardiogram (PCG) recorded heart auscultations between normal and abnormal heartbeat. Bandpass filter is applied to the recorded heartbeat signals to remove noises and other redundant signals which are recorded by the PCG device along with the heartbeat of the patient. The usage of the filter helps in increasing the credibility of the algorithm and give a better accuracy rate in classification of the phonocardiogram (PCG) recorded heart sound signals.

Another contribution of the paper is to divide each heart sound sample into non-overlapping frames and then extract features from them. This will help in improving the accuracy of the developed algorithm. After extracting the features from each frame of the recorded signals, statistical features such as mean and standard deviation were calculated from all frames as a single representative of audio sample. The audio samples are strategically processed to extract informative discriminatory features to determine abnormal and normal heart sounds using supervised classification.

For better classification of the developed algorithm, most prominent features having lower p-value determined from the statistical models of significance were only subjected to classifiers. Another significant contribution of the proposed work is its low time complexity for normal and abnormal heart sound classification. Considering only most discriminatory features for classification improves the performance of classifiers and ensures less time complexity and makes it acceptable for real-time applications.

The remaining structure of this paper work is as follows: Section 2 provides information about the database which has been used to test the developed algorithm. Section 3 describes about the proposed audio signal processing-based methodology for proper classification of the abnormal and normal heart sounds. It includes all the steps, like preprocessing of PCG recordings of heart sounds, feature analysis and selection of discriminatory features using p-values and training of different machine learning algorithms. Section 4 describes the experimental results achieved by the proposed algorithm. Section 4 gives an account about the experimental results obtained from the proposed algorithm and finally, in the last Sect. 5, a discussion about the conclusion of the proposed work is mentioned.

2 Database used: heart sounds—materials

In the proposed methodology, experiments are accomplished on the PCG recorded heart sound signals database which is created by National Institute of Health (National Center for Research Resources) [21]. The PCG recorded heart signals are either from healthy subjects or pathological patients. Each audio belongs to one patient only, and the existence of same patient file is most unlikely in the sets. This database contains both normal as well as abnormal sounds for training and testing purpose.

The heart sounds were recorded from different body locations including four main locations: Aortic area, mitral area, tricuspid area and pulmonic area. These four locations are mostly used by doctors to listen heart sounds. The database is basically categorized into two parts: abnormal and normal heart sound. The abnormal heart sounds were recorded from patient who had a confirmed cardiac disease, typically related to defects of heart valve and coronary artery while the normal sounds were recorded from healthy subjects. Recorded heart sounds of this database belong to both, children as well as adults. All the recorded heart signals are sampled at a frequency of 2000 Hz and are accessible in.wav format. Figure 1 shows two samples of normal/abnormal heart sound recording and their power spectrum. It is clearly observed from the power spectrum diagram that the abnormal heart sound signal has more frequency content as compare to normal heart sound.

Fig. 1
figure 1

a “Normal heart sound” recording b “Abnormal heart sound” recording c Power Spectrum of normal heart sound d Power spectrum of abnormal heart sound

Full size image

3 Proposed methodology: methods

The main aim of the proposed work is to develop a machine learning-based system which can be used for automatic diagnosis of normal and abnormal heart sound using PCG recorded heart beats. The proposed method is explained in Fig. 2 shows the flowchart diagram of the proposed framework for automated diagnosis of cardiac diseases. The proposed framework involves strategic processing and framing of heart sounds to extract informative features followed by an analysis to determine the most prominent features for classification. All extracted features are analyzed based on their p-values, and only those features are subjected to different classifiers which are discriminatory in nature. The main purpose to select the discriminatory features is to obtain a more accurate and precise classification of normal/abnormal heart sounds. The proposed model is categorized into four sub-sections:

Fig. 2
figure 2

Flowchart of the proposed method for diagnosis of cardiac diseases

Full size image

  • Processing and Framing, Extraction of Features

  • Selection of Prominent Features and

  • Supervised classification of normal/abnormal heart sounds.

3.1 Signal preprocessing

Automatic classification of heart sounds in normal and abnormal class is a challenging task due to the insignificant variance in heart sounds that are coming from normal or abnormal subjects. Keeping this in mind, there is a requirement to develop strategic methodology which can properly extract out the prominent and discriminatory features from the heart signals and which may further assist the machine learning-based classifiers in better classification of the heart sound in terms of normal and abnormal heart sound signals. In the proposed work, audio samples of heartbeat are recorded with the help of PCG.

While recording, many noises and other artifacts are also recorded. So, denoising is the first and crucial step for proper analysis of the heart sound. For this purpose, bandpass filter is used. The range of the used bandpass filter is taken from 20 to 500 Hz. Lower cut-off frequency, i.e., 20 Hz is prescribed in the used database. The filtered signal retained all the information values corresponding to the above-mentioned frequency ranges. Other redundant and noisy information above 500 Hz are removed using the bandpass filter. Figure 3a presents the PCG recorded heartbeat signal. It contains different sort of noises which can corrupt the signal during analysis. The signal is passed through a band pass filter and the resultant signal is shown in Fig. 3b.

Fig. 3
figure 3

a PCG recorded heart sound signal b Filtered heart sound signal

Full size image

3.2 Features extraction

For proper classification of the heart sound signal, most prominent features are extracted from the recorded PCG heart sounds. In this developed algorithm, to increase the accuracy of the algorithm, each audio sample is split into non-overlapping frames and then various features are extracted from each frame of audio sample for improving the accuracy of the developed methodology.

Size of audio frame depends upon the size of audio sample in a proportionate optimized manner. All the heart sound signals are sampled at 2000 Hz. Bit rate of the audio samples is 16 bits per sample resulting into 32,000 bps. After extracting the features from each frame of the recorded signals, statistical features such as mean and standard deviation were calculated from all frames as a single representative of audio sample. Figure 4 shows flowchart of the framing of the filtered heart signal.

Fig. 4
figure 4

Flowchart of the method of the framing of the filtered heart sound

Full size image

3.2.1 Extraction of the features from each frame of the heart sound

In this proposed work, six audio features (spectral centroid, zero crossing rate, energy entropy, spectral roll off, volume and spectral flux) are extracted from each frame of audio sample. Then, the statistical parameters are calculated as a single representative of that audio sample which eases the complexity of classification as well as reduces the computational time for classification.


(a) Spectral Centroid It is calculated in terms of weighted mean of frequency in the audio sample using Fourier transform.

$$ {\text{Centroid}} = \frac{{\mathop \sum \nolimits_{n = 0}^{N - 1} f\left( n \right)x\left( n \right)}}{{\mathop \sum \nolimits_{n = 0}^{N - 1} x\left( n \right)}} $$
(1)

where x(n) defines weighted frequency of bin number. f(n) is the center frequency of bin, n is number of bin whose range is 0 to N − 1.


(b) Energy Entropy The heart sound is divided into K sub windows of specific duration. Energy entropy for each frame is calculated by the equation given below:

$$ I_{j} = - \mathop \sum \limits_{i = 1 \ldots k} \sigma_{i}^{2} \log_{2} \sigma_{i}^{2} $$
(2)

where (\( \sigma_{i}^{2} \)) is the normalized energy calculated for each subwindow i.


(c) Spectral Roll off It is expressed as the Kth percentile of the total power spectral distribution in audio signal, where K is 85% or 95% spectral roll off of audio signal. The spectral roll off is also defined as the frequency Ft below which 85% is concentrated magnitude distribution. It can be calculated by the equation given below:

$$ \mathop \sum \limits_{n = 1}^{{R_{t} }} N_{t} \left[ n \right] = 0.85\mathop \sum \limits_{n = 1}^{N} N_{t} \left[ n \right] $$
(3)

where \( N_{t} \left[ n \right] \) is the amplitude of frequency component at frequency bin n and frame t. N is total number of frequency bin.


(d) Spectral Flux Spectral flux calculation is independent of total power and the phase consideration. It can be used for determining the timbre of audio. It can be calculated using the equation given below:

$$ F_{j} = \mathop \sum \limits_{k = 0 \ldots S - 1} \left( {N_{j,k} - N_{j - 1,k} } \right)^{2} $$
(4)

where \( N_{j,k} \): spectral energy of jth frame for Kth sample.


(e) Zero Crossing Rate It is defined as the rate of change of signs along the audio, i.e., it is the rate of change of signal from positive value to negative value or vice versa. ZCR can be calculated as:

$$ Z_{j} = \frac{1}{2S}\mathop \sum \limits_{i = 1 \ldots S} \left| {\text{sgn} \left( {x_{i} } \right) - \text{sgn} \left( {x_{i - 1} } \right)} \right| $$
(5)

3.2.2 Statistical analyses of the extracted features

After extracting the different features frame-wise, statistical features like mean as well as standard deviation of all the extracted features are calculated. This is done so that the calculated statistical parameters act as a single representative of that audio sample which help in reducing the complexity of classification as well as the computational time at the time of classification step.

$$ {\text{Mean}}\,\left( \mu \right) = \frac{{\mathop \sum \nolimits_{r} \mathop \sum \nolimits_{c} \left\{ {J\left( {a,b} \right)} \right\}}}{A \times B} $$
(6)
$$ {\text{Standard}}\,{\text{Deviation}}\,(\sigma ) = \sqrt {\frac{{\mathop \sum \nolimits_{a} \mathop \sum \nolimits_{b} \left\{ {J\left( {a,b} \right) - \mu } \right\}^{2} }}{A \times B}} $$
(7)

3.3 Discriminatory feature analysis and prominent feature selection

In the proposed work, machine learning algorithms are used for automated classification of normal/abnormal heart sounds for diagnosis of cardiac diseases. Accuracy and time complexity of classification algorithm depends upon the discriminatory features that are used in the classification algorithm. If all the extracted features are fed to the classifier, then it may not only affect its accuracy but at the same time may also affect the complexity and CPU time for calculation. Thus, optimized feature selection is a crucial task as it enables the classifier to train faster on the correct subset of features. It also improves the accuracy of the developed model by making it easier to classify between normal and abnormal heart sounds.

Initially, twelve features are extracted from recorded heart sound sample which are further analyzed to find out most prominent features having discriminatory properties for accurate classification of heart sounds. In this proposed methodology, Wilcoxon rank sum test [22] is used which further checks discrimination for all the twelve extracted features. Wilcoxon rank sum test is a statistical hypothesis test which is used to find out the p-value. The most prominent properties are selected on the basis of p-value which is used for classification purpose. When the p-value is lower, then it shows higher discriminatory behavior among the features. Thus, in this paper, the p-values are considered as the standard to select discriminatory features for further classification. Based on the Wilcoxon rank sum rule, p-value is calculated for all 12 extracted features and most prominent features which have lower p-value are finally considered for classification.

Table 2 displays the calculated mean ± standard deviation along with p-value of each of the extracted 12 statistical features from the heart sound database. From Table 1 using p-value, it is quite visible that all features are not showing enough discrimination, which can used for further analysis. In order to reduce the time complexity of the classifiers and to obtain better accuracy, only most prominent features are considered which have lower p-value. These discriminatory features are then fed to different classifiers which are used in the proposed algorithms.

Table 2 Extracted features from heart sounds and their corresponding p-value
Full size table

Table 3 shows the selected 4 prominent features out of the 12 features extracted and their corresponding p-values subjected to classification. Another factor which supported Wilcoxon rank sum (p-value) method to find out discriminating behavior of feature vector is box plot of features. It gives the clear picture of discrimination of feature vector whether they can be used for classification purpose or not. Box plot helps in providing a better view, regarding the extent to which a feature is discriminative in nature.

Table 3 Selected discriminatory features and their corresponding p-value
Full size table

Figure 5 presents the graphs of box plot of the selected prominent features which are further used for the purpose of classification. From the figure, it is clearly observed that the selected features show the enough discrimination which further helps in accurate categorization of abnormal and normal heart sounds using machine learning. Figure 5 shows the graphical distribution of selected features from normal and abnormal heart sounds. It can be seen in Fig. 6 that selected features are more significant to classify abnormal/normal heart sounds and there is a significant gap between the values of the features of two classes under consideration.

Fig. 5
figure 5

Box plots showing discriminating behavior of selected features. a Mean: spectral centroid b Mean: spectral flux c Mean: energy entropy d Mean: zero crossing rate

Full size image
Fig. 6
figure 6

Plots of selected prominent features from normal/abnormal heart sounds (Red Line Plot defines the extracted values of normal heart sounds. Green Line Plot defines extracted values of abnormal heart sounds) a Mean: Spectral Centroid b Mean: Spectral flux c Mean: Energy Entropy d Mean: Zero Crossing Rate

Full size image

3.4 Train different machine learning classifier for the classification of recorded heart sounds

On the basis of Wilcoxon rank sum test, four most prominent features are fed to different classifiers for classifying recorded heart sounds as normal or abnormal class. In this proposed algorithm, different machine learning-based classifiers like Naïve Bayes, support vector machine, random forest and k-nearest neighbor are used. Among these classifiers, the best performing classifier is selected for classification.

Table 4 shows the different classifiers with different parameters used for classification. The main control parameters of the classification algorithms used for the proper classification from recorded heart sounds are explained below:

Table 4 Considered setting parameters for classifiers
Full size table

Kernel function that is used in SVM plays an important for proper training of the SVM. Then, using the grading of the recorded sounds provided by the doctors, support vectors are trained as normal and abnormal. The performance of the algorithm classification is for all possible types of kernel and finally the best parameter is selected for classification using test samples of recorded heart sounds. Figure 7 presents the step wise flowchart of SVM classifier which is used to classify healthy/unhealthy recorded heart sounds.

Fig. 7
figure 7

SVM classification used to classify the recorded heart sounds as healthy or unhealthy heart sound

Full size image

K-Nearest Neighbors algorithm classifies a data point on the basis of its neighbors. It is based on the feature similarity, and its distance measure metrics is selected according to the available training dataset. Choosing the right value of “k” in the processing is known as parameter tuning. It is very important for better accuracy obtained by k-NN algorithm. Figure 8 shows the flow diagram of the k-NN classification algorithm which is used in this work to classify between healthy and unhealthy heart sounds.

Fig. 8
figure 8

Flow of k-NN classifier used to classify the recorded heart sounds as healthy or unhealthy heart sound

Full size image

Naïve Bayes classifier works on Bayes rule and probability measurement. Classifier is trained using training dataset of heart sounds and best parameter setting is selected. On the basis of probability, best prediction is considered for classification of test sample of heart sound. Figure 9 presents flow diagram of Naïve Bayes classification algorithm used for machine learning.

Fig. 9
figure 9

Flow diagram of Naive Bayes classification algorithm

Full size image

In random forest classifier, the main control parameter is the number of decision trees. Using the training data with a set of features and targets, the algorithm of decision tree is generated with some specific set of rules. Performance of the random forest classifier is checked on the same set of rules and can be used to check the performance of the classifier on the testing data. Figure 10 shows the flow diagram of the random forest classifier used in this work.

Fig. 10
figure 10

Flow diagram of random forest classifier algorithm

Full size image

To measure the performance of classifier, different parameters like sensitivity, specificity and accuracy [27] and Equal Error Rate [28] are calculated as follows:

$$ {\text{Sensitivity}}\, = \,\frac{\text{TP}}{\text{TP + FN}} $$
(8)
$$ {\text{Specificity}}\, = \,\frac{\text{TN}}{{{\text{TN}} + {\text{FP}}}} $$
(9)
$$ {\text{Accuracy}}\, = \,\frac{{{\text{TP}} + {\text{TN}}}}{{{\text{TP}} + {\text{FP}} + {\text{TN}} + {\text{FN}}}} $$
(10)
$$ {\text{Equal}}\,{\text{Error}}\,{\text{Rate}}\,\left( {\text{ERR}} \right)\, = \, \frac{{{\text{FP}} + {\text{FN}}}}{{{\text{TP}} + {\text{FP}} + {\text{TN}} + {\text{FN}}}} $$
(11)

where TP is defined as true positive which presents total number of recorded heart samples from abnormal category which are classified as abnormal, TN is defined as True Negative which means the total number of recorded heart samples from normal category which are classified as normal, FP means False Positive which is used to expressed the total number normal heart samples which are classified as abnormal while FN means False negative which is used to defined the total number of abnormal heart samples which are classified as normal.

The performance of classifier is measured at various possible parameters, and best classification performance is considered for the classification of heart sounds.

4 Experimental results and discussion

In the proposed methodology, the experiments are accomplished on a total of 500 heart sounds recorded from healthy patients as well as from those patients who are suffering from some sort of cardiac diseases. No overlapping audio sample is used in the experiment for training and testing purpose. Table 5 gives a detail information about the distribution of 250 Normal and 250 abnormal heart sounds which are used to train the classifiers and for testing of the developed algorithms.

Table 5 Heart sounds database used for experiments
Full size table

Table 6 presents the performance of different classifiers which are used to check the performance of the trained algorithm. Table shown is divided into two sections: first one describes the performance of the classifiers without feature selection, while the next shows the performance when the prominent feature are selected. From the table, it is clearly visible that all four classifiers achieved more than 90% accuracy for classification of PCG recorded heart sounds as normal and abnormal. The best accuracy of 97.78% is provided by the SVM classifier with specificity of 97.78%, sensitivity of 98.46% and equal error rate (EER) of 2.22% for the diagnosis of cardiac disease which can be considered as potentially an important contribution for screening of cardiac diseases.

Table 6 Comparative performance of different classifiers for normal/abnormal heart sounds with/without feature selection
Full size table

Tenfold cross-validation is performance to check the accuracy of each of the used classifier. In tenfold cross-validation, the dataset is nearly divided into k-equal segments which is called folds. Then ten iterations of training and testing are done in such a way that in each iteration one-fold is held for testing purpose while the remaining ninefolds are used for training. In this algorithm, the data are arranged in such a way that in each fold classes of both heart sound samples are comprised equally. Table 7 presents tenfold cross-validation is performance to check the accuracy of each of the used classifier. Accuracy of each classifier is calculated by taking the average of the total accuracy of tenfolds. In tenfold cross-validation, the best classifier is SVM with the accuracy of 95.62%. On evaluating or comparing all performed classifiers, classifiers with prominent features are taken into consideration as they are providing best results.

Table 7 Comparative performance of tenfold cross-validation of different classifiers for normal/abnormal heart sounds
Full size table

Figure 11 presents the receiver operating characteristic (ROC) curve for all the used classifiers that are SVM, random forest, Naïve Bayes and k-NN, in a single plot. It can be seen in Fig. 10 that for all four classifiers, ROC curve is near to (0, 1) coordinate of the ROC space which is the indication of best possible prediction.

Fig. 11
figure 11

ROC curve for random forest, Naïve Bayes, SVM and k-NN classifier

Full size image

It is also clear in Fig. 11 that there is no significant difference in ROC curve for all used classifiers while classifying the normal and abnormal heart sounds. The area under the ROC curve or “AUC” (Area under curve) was calculated from ROC curve for all four classifiers. The value of AUC was 0.9827 for SVM classifier, 0.9956 for k-NN classifier, 0.9870 for Naïve Bayes classifier and 0.9912 for random forest classifier which shows the perfect classification. The EER is the point on the ROC curve where negative and positive rates are equal. While developing a system, this value is kept as small as possible.

Figure 12 is the graphical representation of comparative performance of classifiers in terms of accuracy with/without prominent feature selection and using tenfold cross-validation. Similarly, Fig. 13 presents the comparative performance of all the used classifiers in terms of time complexity with/without feature selection and using tenfold cross-validation.

Fig. 12
figure 12

Graphical representation of comparative classifier performance in terms of accuracy with/without feature selection and tenfold cross-validation

Full size image
Fig. 13
figure 13

Graphical representation of comparative execution time in terms of accuracy with/without feature selection

Full size image

The performance of different classifiers without feature selection, with prominent feature selection and with tenfold cross-validation is shown in Table 6. It clearly indicates that performance of machine learning algorithms for automated diagnosis of cardiac diseases has been improved with prominent feature selection method. The SVM, Naïve Bayes and k-NN classifiers show the best performance in terms of more accuracy and less time complexity with prominent feature selection method. All four classifiers with feature selection method achieved accuracy more than 95% for classification of abnormal and normal heart sound classification which can be considered as encouraging.

The classification algorithms are implemented on a workstation with computing tool MATLAB R2013a installed. The workstation has a configuration of Intel® Core™ i5-4210U CPU@ 1.7 GHz processor, 4 GB RAM, 64-bit operating system, × 64-based processor. The proposed automated system required less than 2.45 s time to test single sample which indicates that the proposed model can be considered as a significant contribution for diagnosis of cardiac diseases in real time.

5 Conclusion

In this developed algorithm, a supervised machine learning-based classification technique is applied to classify the PCG recorded heart sounds as normal/abnormal. Bandpass filter is used to remove the noise that gets recorded along with the heart sound. Presence of any sort of noise creates difficulty in extracting the useful information from the recorded signals of heart sound. Various statistical features are extracted from filtered heart sound signals. Wilcoxon rank sum is employed to find out the most prominent and discriminatory features among the extracted features. Through Wilcoxon rank sum, p-value of all the extracted features is calculated. In this paper, p-value is considered as an important parameter, which is used to filter out the discriminatory features for further analysis. This reduction in generalized features space to a more specific space will ensure better and more efficient classification. When the p-value is lower, it shows higher level of discriminatory behavior among the features. In the proposed work, the features were strategically considered which are likely to be affected during an abnormal heart functioning. Selection of most prominent features improved the performance of machine learning algorithms in terms of time complexity and accuracy. The proposed method achieved 97.78% accuracy with 97.78% sensitivity and 97.78% specificity for abnormal and normal heart sound classification when tested on heart sound database. The proposed model of automated diagnosis shows high accuracy with low time complexity as compared to existing method of classification. Future work will be focused on the detection of various kind of the abnormal heart diseases. The algorithm can improve by detecting other body sounds such as sounds coming from lungs.