Table 3

Tabular summarisation of search on the use of machine learning algorithms in rhythm classification during CPR

Study	Number of ECG segments used	Study design	Algorithm used	Sensitivity	Specificity	Accuracy	Limitations
Jekova et al26	1545	End-to-end analysis of ECG during CPR in OHCA using CNN	CNN	89.0%	91.7%	–	Data did not contain statistically significant numbers of shockable VT
Hajeb-Mahammadalipour et al51	23816	Development of an automated condition-based filter to removed CPR artefacts for accurate rhythm analysis during CPR	Condition based filtering algorithm followed by ANN	94.5%	88.3%	89.2%	Assumed constant rate of chest compressions constant within the 14 s period Difficulty removing artefacts from asystole ECGs and lack of sufficient asystole ECGs in training set
Hajeb-Mahmmadalipour et al52	3872	Analysis of ECG rhythms superimposed with CPR artefacts using a CNN	CNN	95.2%	86.0%	88.1%	Artificially introduced artefacts from AEDs in asystole not real-life traces Not tested during asystole
Didon et al25	2916	To present new combination of algorithms for rhythm analysis during CPR in AED	Analyse While Compressing (AWC)	92.10%	>99%	–	Small sample of VT rhythms Still requires 'hands-off' reconfirmation of classification in 34.4% of cases
Isasi et al23	272	Rhythm classification during CPR using a recursive least squares filter followed by CNN	Recursive least squares filter followed by CNN	95.8%	96.1%	96.0%	Recursive least squares filter requires thoracic impedance to remove ECG artefacts
Hu et al53	1578	Two-step analysis of ECG during chest compressions whereby if shockable rhythm not identified, chest compression-free analysis occurs	A two-step analysis through CPR algorithm	93.60%	99.50%	–	Small sample size of coarse VT The OHCA cardiac arrests were not treated with a defibrillator until they arrived at hospital Short ECG segments
Isasi et al54	2203	Use of machine learning algorithms following CPR artefact filtering for reliable shock decisions	Least mean squares filter followed by ANN, SVM, Kernel Logistic Regression or Random Forest classifier	94.5%	95.5%	96.0%	Computer based study not ‘bench’ simulation study
Fumagalli et al55	2701	Analysis of ECG during chest compressions with 3 s pause to re-confirm rhythm	Analysis During Compressions with Fast Reconfirmation (ADC-FR) algorithm	95.0%	99.0%	–	Requires thoracic impedance for removal of ECG artefact
Yu et al24	1017	An adaptive filter which can eliminate CPR artefacts from corrupted ECGs without any reference channels can be used for non-shockable rhythm detection during CPR	ANN	95.0%	80.0%	–	Tested with artificial mixtures of clean human ECGs and CPR artefacts collected from pigs Only 24 CPR artefacts produced and superimposed onto the ECG segments

ANN, artificial neural network; CNN, convolutional neural network; OHCA, out of hospital cardiac arrest; VT, ventricular tachycardia.