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

Nguyen, Minh Tuan; Nguyen, Thu-Hang T.; Le, Hai-Chau

doi:10.1186/s12938-022-00993-w

BioMedical Engineering OnLine

Table 3 Summary of intelligent ML-based SAA using alternative signals

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
[16], 2016	CUDB, MITDB, AHA	Subsignals using DWT, 3 s	Use of DWT for reconstruction of subsignals.	1	NA	Five-folds CV	Insignificant improvement for classification performance.	Only a feature and SVM classifier.
			Calculation of the number of samples which is larger or smaller than positive or negative thresholds during 1 s segment.				Time consumed for FE may be over segment length of 3 s.	Relative short of segment length.
			Use of average numbers of samples as a feature for SVM classifier.
[27], 2017	MITDB	Subsignals using DWT, 5.7 s	5 levels of wavelet coefficients using DWT.	20	NA	NA	Limited database.	Peak extraction from wavelet coefficients.
			Peak extraction from wavelet coefficients, plotted in 3D PRS.				No FS and validation.
			NEWFM classifier using 20 features considered as distances between origin of coordinates axis and peaks.
[40], 2017	CUDB, VFDB, MITDB	Subsignals using DWT, 10 s	DWT with 4 levels-decomposition.	31	SFFS	Five-folds CV	Only 1 classifier.	FFC of 10 features.
			Feature extracted from wavelet coefficients.				No validation performance for all ICFs.	The best ranking method of ReliefF.
			SFFS to select 14 features.
			Feature ranking using 6 methods for set of 14 features.
			KNN classifier using different sets of features of 6 ranking methods.
[38], 2019	CUDB VFDB, MITDB	Subsignals using DWT, 5 s	Performance comparision of C4.5 and SVM for detection of VF, VT.	24	GRAE	10-folds CV	Highest performance of all ICF.	Generation of signals concentration on VT and VF components based on DWT.
			Using DWT as low-pass and high-pass filters for generation of alternative signals.				Ineffective FS.
			Features extracted from alternative signals.
[39], 2016	CUDB, VFDB, MITDB	Subsignal using wavelet decomposition, 2 s	Analysis on wavelet decomposition to design an optimal low-pass filter showing a minimum stopband ripple energy.	12	NA	10-folds CV	Limited number of ICFs.	Selection of six subsignals based on orthogonal conditions.
							No FS.	Productive SVM for SH rhythm detection.
								Relative short of segment length
[41], 2016	CUDB, VFDB, MITDB	Modes using VMD, 5 s	5 modes using VMD.	9	FS based feature scoring	Five-folds CV	Limited number of ICFs.	Modes using VMD for FE.
			FE from first 3 modes.				Hand-picked data.	FFC of 7 features.
			The FS based feature scoring to select an FFC of 7 features.				Random reconstruction of modes.
			Validation of the FFC using RF and five-folds CV.
[43], 2017	AFDB MITDB, NSRDB	Modes using VMD, 8 s	Decomposition of ECG into 5 modes.	20	NA	Five-folds CV	No FS.	Effective entropy features.
			Sample entropy and distribution entropy of modes.				Hand-picked data.	High performance of SVM with KBF kernel among others.
			Performance of 2 ML classifiers for normal, AF, and VF scenario.				Random generation of modes.
							Limited number of ICFs
[42], 2018	CUDB, VFDB, MITDB	Modes using adaptive VMD, 5 s	5 modes using adaptive VMD.	30	NA	10-folds CV	No FS.	Optimal parameters for adaptive VMD.
[42], 2018	CUDB, VFDB, MITDB	Modes using adaptive VMD, 5 s	10-folds CV for Boosted CART using all ICFs.	30	NA	10-folds CV	Simple selection of VMD parameters.	Optimal parameters for adaptive VMD.
[28], 2018	CUDB, VFDB,	Modes using dimensional Taylor Fourier transform, 8 s	Decomposition of ECG segment into oscillatory modes using dimensional Taylor Fourier transform.	20	NA	NA	Low performance.	New diagnostic features of magnitude and phase differences using dimensional Taylor Fourier transform.
			20-dimension feature vector based on magnitude and phase differences.				No FS and FV.
			LSSVM classifier for detection of shock/non-shock, VT/VF, and VF/non-VF.				Only 1 classifier.
[17], 2009	MITDB	Intrinsic mode functions using EMD, 7 s	Use of intrinsic mode function with EMD.	2	NA	NA	No validation.	Orthogonality of IMFs as the features.
[17], 2009	MITDB	Intrinsic mode functions using EMD, 7 s	Calculation of 2 angles between first 3 IMFs for Bayer decision theory.	2	NA	NA	Limited database	Orthogonality of IMFs as the features.
[18], 2017	VFDB AHA	Image of time-frequency, 150 ms	Construction of time-frequency image.	1	NA	NA	Only 1 feature.	Algorithm design for multiple classification using different binary ML classifiers.
			Performance comparison of different ML classifiers for classification of normal, VF, VT, and other rhythms.				No validation.
							Complexity due to 3 ML classifiers for multiple classification
[19], 2018	VFDB AHA	Time-frequency representation image, 1.2 s	Extraction of image using Hilbert transform and Time-frequency representation techniques.	1	NA	Five-folds CV	Only 1 feature.	Effective feature of TFRI image.
[19], 2018	VFDB AHA	Time-frequency representation image, 1.2 s	Use of multiple ML classifiers to detect normal, VF, VT, other rhythms.	1	NA	Five-folds CV	Increase in complexity due to binary algorithms for multiple classification	Hierarchical topology of 3 ML classifiers.

Back to article page

ISSN: 1475-925X

Contact us

Submission enquiries: journalsubmissions@springernature.com