Skip to main content

Table 2 Summary of intelligent ML-based SAA using stand-alone ECG

From: A review of progress and an advanced method for shock advice algorithms in automated external defibrillators

Refs., Year Database Signal for FE, Segment length Method description ICF FS FV Limitation Key Findings
[20], 2007 CUDB, VFDB, AHA Stand-alone ECG, 10 s Use of 10 ICFs as the input of Linear discriminant analysis to select an FFC of 4 features. 10 Linear discriminant analysis embedding FS NA Limited number of ICFs. FFC of 4 features.
No validation. Prediction of success of defibrillation.
[25], 2011 MITDB Stand-alone ECG, QRS comlex of 200 points Performance comparison of KNN, neural networks and ensemble based methods. 7 NA NA Limited database. Better performance of DECORATE model than others.
Limited number of ICFs.
No FS and FV
[21], 2014 CUDB, VFDB, MITDB Stand-alone ECG, 8 s SBFS including SVM and boots trapping to select an FFC of 7 features on training data. 13 SBFS using SVM and boostrapping NA Limited number of ICFs. FFC of 9 features.
  Performance of SVM using a FFC on testing data. No validation.
[22], 2012 CUDB, VFDB, MITDB   Extraction of 11 Vleak features. 11 NA NA Limited number of ICFs. FFC of 11 features and SVM.
  Comparison performance of SVM and VLeak threshold for VF/non-VF and shock/non-shock. No FS and FV. Better performance of SVM than Vleak.
[23], 2014 CUDB   Use of Hilbert transforms for peak extraction, phase space reconstruction, time domain analysis. 15 NEWFM embedding FS. NA Limited database. FFC of 11 features.
  NEWFM embedding FS to select an FFC of 11 features. No separated data for FS and testing.
   No validation.
[29], 2016 CUDB, VFDB, AHA, OHCA   SBFS including 2 ML classifiers and bootstrapping to select 2 CFCs. 30 SBFS using Logistic regression, Boosting and boostrapping. Bootst-rapping Lower validation performance of CFCs than all ICFs. Large number of ICFs.
  Validation of CFCs and a combination of all ICFs using 5 ML classifiers and bootstrapping. OHCA data requires two times more features than public data.
[30], 2017 CUDB, VFDB, MITDB   SVM to rank 26 ICFs and selection of 19 features. 26 Feature ranking with SVM Record-based data division. No validation for all ICFs. FFC of 3 features
  Validation of every combination of 19 features using SVM and random data division. Database-based data division.
[24], 2018 CUDB, VFDB, MITDB, AFDB   Algorithm design for classification of VF/non-VF, Atrial fibrillation/non 6 NA NA Limited number of ICFs. Effective features computed from time-delay algorithm.
  -Atrial fibrillation, premature ventricular contraction/non-premature ventricular contraction, and sinus arrhythmia. No FS and FV.
  SVM and Bayer decision tree for VF/non-VF classification.  
[31], 2016 CUDB, VFDB, MITDB   Use of RF and 10folds CV to validate the combination of all ICFs for different window lengths. 17 NA 10-folds CV No FS Best performance for overlapping 8 s-segment.
[32], 2014 CUDB, VFDB, MITDB Stand-alone ECG, 5 s GA based feature ranking. 14 GA five-folds CV -Limited number of ICFs. FFC of 2 features.
  Performance investigation of every combination of 9 features using SVM. No validation for all ICFs.
  Validation of 9 combinations using five-folds CV and SVM.  
[33], 2018 CUDB, VFDB   GA based feature ranking for selection of 7 good features. 11 GA Five-folds CV Limited number of ICFs. FFC of 4 features.
  Performance estimation of SVM using every feature combination of good features. Only 1 classifier
  Validation performance of SVM using 6 combinations with five-folds CV. Similar method of [15]
[34], 2018 CUDB, VFDB   C4.5 for classification of normal, VF, and VT segments 13 GRAE 20-folds CV Highest performance of all ICFs Identification of confidence factor value of C4.5
  Feature ranking using gain ratio attribute evaluation.
  Investigation of different confidence factor for C4.5
[37], 2021 MITDB, CUDB,VFDB   Application of SVM and AdaBoost for the FS based differential evolution algorithm and classification of VF and non-VF rhythms. 17 Differential evolution algorithm 10-folds CV Only SVM. Effective AdaBoost for data weight assignment to improve SVM classification performance.
  Extraction of 17 conventional features and selection of 3 as the optimal feature subset Limited number of ICFs.
   No validation for all ICFs.
[35], 2012 VFDB, AHA Stand-alone ECG, 1.024 s Extraction of temporal, spectral, and time-frequency features. 37 SVM-bootstrap resampling SVM-recursive feature elimination Filter methods 5-folds CV Performance analysis based on 1 ML classifier. Highest performance of FFC of 3 features.
  Comparison of SVM-bootstrap resampling, SVM-recursive feature elimination and filter methods. Only 1 s-segment Efficient SVM-bootstrap resampling.
[26], 2012 VFDB,AHA   Extraction of temporal, spectral, and time-frequency features. 27 Bootstrap resampling NA Performance analysis based on a ML classifier. Self organizing map using an FFC of 11 features.
  Selection of 11 features using bootstrap resampling based feature selection.
  Self organising map for classification.
[36], 2019 CUDB, VFDB, NSRDB Stand-alone ECG, 3 s Investigation of 47 time domain and wavelet features. Selection of 25 by FS. 47 Gaussian GA 3-folds CV Investigation of a classifier. 17 wavelet features (out of 25) show the significant efficiency of wavelet method.
Detection of VF/VT and normal ECG segment by the first SVM classifier. Validation on only a database. Practical application for AED processor.
Discrimination between VF and VT by the second SVM classifier. Low average CV performance Reletive short of segment length.
Hardware implementation of the proposed algorithm for the AED.