A Review of Computer-Aided Heart Sound Detection Techniques

Table 3

Feature extraction and classification methods of PCG signals.

Year	Author	Feature extraction methods	Classifier	Database	Result
2019	Shi et al. [33]	Feature extraction algorithm of Springer	AdaBoost	PhysioNet and PASCAL	ACC: 96.36%
2019	Nogueira et al. [34]	MFCC	SVM	PhysioNet	Sensitivity	Specificity	Accuracy
					91.87%	82.05%	97%
2019	Cheng (without segmentation) [35]	Envelope autocorrelation	SVM	HSCT11 dataset	Accuracy all could reach to 100%
2018	Meintjes et al. [36]	CWT	SVM, kNN	PhysioNet	MAcc: 86%
2018	Hamidi et al. [37]	Curve fitting, MFCC	Euclidean distance	Dataset A from PhysioNet	MAcc: 92%
				Dataset B from PhysioNet	MAcc: 81%
				Dataset C from PhysioNet	MAcc: 98%
2018	Juniati et al. [38]	DWT	kNN, Fuzzy c-means clustering	40 normal heart sounds, 40 extra systole, 40 murmurs	MAcc: 86.17%
2017	Kay et al. [39]	CWT, MFCC	BP neural networks	PhysioNet	MAcc: 85.2%
2017	Karar et al. [40]	DWT	Rule-based classification tree	22 sets of heart sounds and noise data from the public database of the CliniSurf medical school	MAcc: 95.5%
2017	Zhang et al. [41]	Tensor decomposition	SVM	Dataset A: normal heart sounds, extra systole, murmurs, artificial heart sounds	MAcc: 76%
				Dataset B: normal heart sounds, extra systole, murmurs	MAcc: 83%
				Dataset C: normal heart sounds, abnormal heart sounds	MAcc: 88%
2017	Langley and Murray (without segmentation) [42]	/	Wavelet entropy	PhysioNet	Sensitivity	Specificity	Accuracy
					94%	65%	80%
2017	Whitaker et al. [43]	Sparse coding	SVM	PhysioNet	Sensitivity	Specificity	MAcc
					84.3%	77.2%	80.7%
2017	Li et al. [44]	FFT	BP neural networks	PhysioNet	Sensitivity	Specificity	MAcc
					68.36%	94.01%	88.56%
			Logistic regression		Sensitivity	Specificity	MAcc
					75.68%	87.71%	72.56%
2016	Deng and Han (without segmentation) [45]	DWT	SVM-DM	Dataset A from PASCAL	The highest total precision of 3.17
				Dataset B from PASCAL	The highest total precision of 2.03
2015	Zheng et al. [46]	EMD	SVM	A dataset collected from the healthy volunteers and CHF patients	Sensitivity	Specificity	Accuracy
					96.59%	93.75%	95.39%
2015	Safara [47]	Wavelet packet tree	Higher-order cumulants (HOC)	A set of 59 heart sounds from different categories: normal heart sounds, mitral regurgitation, aortic stenosis, and aortic regurgitation.	Best classification accuracies: 99.39%
2011	Yuenyong et al. (without segmentation) [48]	DWT	Neural network	Several on-line databases and recorded with an electronic stethoscope	Tenfold cross-validation: 0.92 for noise free case, 0.90 under white noise with 10 dB signal-to-noise ratio (SNR), and 0.90 under impulse noise up to 0.3 s duration