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 |
---|---|---|---|---|---|---|---|---|---|
Nogueira et al. [38] | 2017 | – | Springer’s segmentation algorithm [15] | Classification of heat images generated from MFCC and time features using SVM, KNN, CNN, and random forest | Normal and abnormal heart sound | PhysioNet CinC challenge, 2016 | 13404 heart image | Sensitivity, specificity, and overall score | 96.47, 72.65, and 84.56% |
Chen et al. [36] | 2017 | MFCC | K-means and DNN | K- means and DNN | Normal heart sound | Private dataset | 920 beats | Precision, recall, and F measure | All measures > 91% |
Potes et al. [32] | 2016 | Signal decimation and low pass filtering | Time, spectrum, MFCC-based features, CNN, AdaBoost-Abstain classifier, and ensemble classifier | Time, spectrum, MFCC-based features, CNN, AdaBoost-Abstain classifier, and ensemble classifier | Normal and abnormal heart sounds | PhysioNet CinC challenge, 2016 | 3240 beats and a blind dataset | specificity, sensitivity, and overall score | 92.42, 77.81, and 86.02% |
Ortiz et al. [39] | 2016 | – | Springer et al. [15] | SVM using time, MFCC, DWT, and wavelet | Normal and abnormal heart sounds | PhysioNet CinC challenge, 2016 | 3240 beats and a blind dataset | Test score | 82.4% |
Kotb et al. [17] | 2016 | Pre-emphasis filtering and windowing | MFCC | Heart murmur model using HMM | Normal and abnormal heart sounds | Private dataset | 1069 normal and abnormal heart sound | Sensitivity, correct classification rate | 98 and 96% |
Tang et al. [40] | 2016 | Bandpass filtering | Springer et al. [15] | ANN using time, kurtosis, MFCC, energy, power spectrum cyclostationary features | Normal and abnormal heart sounds | PhysioNet CinC challenge, 2016 | 3240 beats and a blind dataset | Sensitivity, specificity, and over all score | 80.7–81.2%, 82.9–85%, and 81.8–83.6% |
Rubin et al. [41] | 2016 | – | Springer et al. [15] | CNN using heat maps generated using MFCC features | Normal and abnormal heart sounds | PhysioNet CinC challenge, 2016 | 3240 beats and a blind dataset | Specificity, sensitivity, and overall score | 95, 73, and 84% |
Redlarski et al. [35] | 2014 | – | Modified linear predictive coding(LPC) | ANN, SVM-MCS | Normal, abnormal heart sounds, and clicks | 3M Poland microphone samples | 72 signals | Accuracy | 66–100% |
Naseri et al. [29] | 2013 | – | Frequency and energy features | Frequency and energy features | Normal and noisy heart sound segments | Signals recorded at Rajaei Cardiovascular, Medical and Research Center, Iran | 63 clean and 63 noisy segments | Sensitivity and positive predictive value | 91.24 and 92.86% |
Singh et al. [30] | 2013 | – | Time domain, frequency domain, cepstrum, and statistical features | Time domain, frequency domain, cepstrum, statistical features, Bayes Net, naive Bayes, GSD, and Logit Boost | Normal heart sounds and murmurs | PASCAL CHSC2011 data set A and B | 30 normal and 30 murmurs | Specificity, sensitivity, and accuracy | 93.3, 93.3, and 93.3% |
Ari et al. [33] | 2010 | Low pass filtering, normalization, and Shannon energy envelop calculation | Low pass filtering, normalization, Shannon energy envelop calculation, and wavelet decomposition | Least square SVM with LMS algorithm | Normal heart sounds, aortic insufficiency, aortic stenosis, atrial septal defect, mitral regurgitation, and mitral Stenosis | Data recorded at Maulana Azad Medical Institute, Delhi, India | 512 cycles from 64 signals | Classification rate | 86–100% |
Chauhan et. al. [16] | 2008 | Low pass filtering, normalization, and Shannon energy envelope calculation | Time duration and MFCC | Time duration, MFCC, and HMM | Normal heart sounds, snaps, murmurs, and clicks | Private dataset | 41 Samples | Correct/ incorrect recognition rate | 80–100% |
Ali et al. [34] | 2007 | Low pass filtering | Shannon energy envelop withassification technique adaptive window and heart beat modeling | Time duration in Shannon energy envelope | Normal heart sounds, murmurs, and clicks | E-General Medical | – | – | – |
Liang et al. [31] | 1997 | Signal decimation, low pass filtering and Shannon envelop formation | Shannon energy envelope with adaptive window, heart beat modeling, and time- based peak detection of Shannon energy envelope | Time-based peakdetection of Shannon energy envelope | Normal and abnormal heart sounds | Private dataset | 37 recording composed of 515 cycles | Single cardiac cycle correct/ incorrect detection | 93% |
Yogana-than et. al. [37] | 1976 | – | FFT | FFT | Normal heart sounds | Private dataset | 29 signals | – | – |