From: Localization and classification of heart beats in phonocardiography signals —a comprehensive review
Study | Year | Pre-processing techniques | Localization techniques | Classification techniques | Beat types | Data set | Signals statistics | Evaluation metrics | Results |
---|---|---|---|---|---|---|---|---|---|
Kay et al. [73] | 2017 | – | Springer’s segmentation algorithm | ANN using beat duration, MFCC, CWT, and complexity features | Normal and abonormal heart sounds | 2016 PhysioNet challenge | – | Accuracy | 85.2% |
Nabih et al. [77] | 2017 | Reconstruction based on DWT and IDWT | Db 10 wavelet | ANN | Normal heart sounds | PASCAL CHSC 2011 Dataset A and B | 170 | SNR, RMSE, and PRD | – |
Jain et al. [69] | 2017 | Adaptive method for DWT coefficient reduction for denoising PCG | – | – | Normal heart sounds | E-General Medical and 10 signal privately recorded | – | % detection of S1 and S2 | 55.1–100% |
Mostafa et al. [71] | 2016 | Downsampling, bandpass filtering, and Schmidt spike removal algorithm [104] | Springer’s segmentation algorithm [15] | FDA, ANN, and voting- based classification | Normal and abnormal heart sounds | 2016 PhysioNet challenge | 4430 signals | Specificity, sensitivity, and over all score | 87.2–97%, 81.1–96%, and 84.1–98% |
Jain et al. [68] | 2016 | Adaptive thresholding of DWT coefficient reduction for denoising PCG | – | – | Normal heart sounds | Private dataset composed of simulated and real life data | – | % detection of S1 and S2 | 86.6–100% |
Boussaa et al. [72] | 2016 | – | – | ANN-based classification using MFCC and DWT features | Normal heart sounds and murmurs | MIT BIH [103] | 50 signals from MIT BIH dataset | Correct decision rate | 100% |
Zahhad et al. [67] | 2016 | DWT-based denoising | – | Classification based on MFCC, LFCC, WPCC, and NLFCC features | Normal heart sounds and murmurs | BioSec. PCG dataset [105] and PASCAL CHSC 2011 dataset A and B | 21 from BioSec. PCG dataset and 206 signals from PASCAL CHSC 2011 | Correct recognition rate | – |
Pedrosa et al. [62] | 2014 | Morlet mother-based continuous wavelet denoising | Time domain, time-frequency domain, perceptual analysis, non-linear, and chaos-based analysis | Time domain, time- frequency domain, perceptual analysis, non-linear, and chaos-based analysis | Normal heart sounds and murmurs | Data recorded at Real Hospital Portugues in Recife, Brazil and PASCAL CHSC 2011 dataset A and B | 72 Digiscope signals and 111 signals from PASCAL CHSC 2011 | For segmentation: sensitivity and positive predictive, and for murmur: sensitivity and specificity | 89.2 and 98.6%; and 52.38–69.67%; and of 46.91–79.40% |
Zheng et al. [63] | 2013 | Normalization, DWT and envelogram | MFCC-based feature extraction | MFCC-based feature extraction | Normal heart sounds and heart sounds with pathological symptoms | E-General Medical | 60 signals | Segmentation rate | – |
Marques et al. [64] | 2013 | – | Stationary wavelet transform to segment the signal | Hierarchical clustering to distinguish S1 and S2 | Normal heart sounds | PASCAL CHSC 2011 | 30 signals from dataset A and 120 signals from dataset B | Error reduction | – |
Yiqi Deng, Peter J Bentley [65] | 2012 | Decompose and reconstruct using wavelet decomposition using 4 level Db6 filters and refilter the signal using spectrogram | Time duration between peaks and peak amplitude | Time duration between peaks and peak amplitude | Normal heart sounds, murmurs, extra sounds and artifact | PASCAL CHSC 2011 | Dataset A and dataset B | Precision, sensitivity, and specificity | 77.67 and 45.83%, 43.75–50.85%, and 44.44–58.2% |
Naseri et al. [61] | 2012 | baseline removal, wavelet base noise removal, and normalization | Time duration, peak value, and time duration between peaks | Time duration, peak value, and time duration between peaks | Normal heart sounds, murmurs, and scouffles | Rajaei Cardi- ovascular, Medical and Research Center, Iran | 50 signals of total duration of 52 min | Sensitivity and positive predictive value | 99.00 and 98.60% |
Song et al. [55] | 2012 | LMS and Db3 wavelet | Time durations extracted from normalized, Shannon envelop formation | Peak and duration based, fuzzy neural network with structure learning (FNNSL) | Normal and abnormal heart sounds | – | 14 signals composed of private and public subsets | Classification rate | 100% for normal and 50% for abnormal |
Zhihai Tu et. al. [59] | 2010 | Db6 wavelet | Heart beat time duration | Hilbert transform-based classification | Normal heart sounds | Shandong University database and real-time device | 123 recordings from database and 100 from real-time device | Correct segmentation rate | 94.60% |
Babaei et al. [56] | 2009 | – | Db4 and inverse Db4 | Db4 and inverse Db4, ANN and statistical classifier | Normal heart sounds, aortic insufficiency, aortic stenosis, and pulmonary stenosis | Tehran Heart Center, Iran | 372 cycles | Accuracy | 94.42% |
Gupta et al. [60] | 2007 | Signal decimation, low pass filtering and signal normalization | Homomorphic filtering, time duration and K-means | Wavelet Db2-based features, GAL, and MLP-BP | Normal heart sounds, systolic murmur, and diastolic murmur | Singapore General, Hospital, Singapore | 41 signals composed of 340 cycles | Classification rate | 90.29% |
Gupta et al. [66] | 2005 | Signal normalization | Peak detection based on homomorphic filtering and K-means | Db2 wavelet-based features and MLP-BP | Normal heart sounds, systolic murmur, and diastolic murmur | Data recorded at Singapore General Hospital, Singapore | 340 cycles | Classification and segmentation accuracy | 94.5 and 90.45% |
Liung et al. [58] | 1998 | – | wavelet packet decomposition | Wavelet packet decomposition | Pathological and physiological murmurs | – | 85 signal of 7–12 s | Accuracy | 85% |
Huiying et al. [57] | 1998 | Decompose and reconstruct signal using Daubechies wavelet | Shannon energy envelop-based S1 and S2 location | Feature vector based on systole time duration, diastole time duration, and ANN | Normal heart sounds and murmurs | Private dataset | 78 recordings | Accuracy | 74.4% |