Table 6.

Comparison of existing algorithms for detection of shockable rhythms.

Reference	Brief approach	Dataset used	Accuracy reported
Figuera et al. (2016)	An ML-algorithms with built-in feature selection capabilities were used to determine the optimal feature subsets for classification. Patient-wise bootstrap techniques were used to evaluate algorithm performance on public database	Validated on the VFDB and the CUDB datasets	Sensitivity = 96.6%, Specificity = 98.8%
Kwon et al. (2018)	The authors proposed an embedded microcontroller where an ECG sensor is used to capture, filter and process data, run a real-time VF detection algorithms developed a VF detection algorithm, via Time Delay (TD), based on phase space reconstruction.	Open access MIT-BIH dataset	Sensitivity = 96.56%, Specificity = 81.53%
Krasteva et al. (2020)	A deep convolutional network was proposed and studied on Holter ECG recordings for detection of shockable and non-shockable rhythms. The impact of various network hyper-parameter tuning was reported	The data used in the study contains a wide variety of non-shockable and shockable rhythms from two sources: public Holter ECG databases from continuously monitored patients with ventricular arrhythmias, and OHCA databases recorded by AEDs from patients in cardiac arrest.	For analysis on short windows (2 s): Sensitivity 97.6% =, Specificity = 98.7%. For analysis on long windows (5 s) : Sensitivity = 99.6 % Specificity = 99.4 %
Jeon et al. (2020)	A deep architecture comprising convolutional layers and recurrent networks for classification of ECG beats. Furthermore, a lightweight model is proposed with fused RNN for speeding up the prediction time on central processing units (CPUs)	The authors used 48 ECGs from the open access MIT-BIH Arrhythmia Database, and 76 ECGs were collected with S-Patch devices developed by Samsung SDS	For the baseline model: Sensitivity = 99.86%, Specificity = 98.31% for the light-weight model: Sensitivity = 99.92%, Specificity = 99.11%
Our proposed approach	A CNN-LSTM architecture is proposed for classification of VF, VT and other rhythms from ECG	The approach is evaluated on CUDB and VFDB datasets	Detection rate of shockable rhythms (VF and VT) on CUDB: very small windows (2 s) Sensitivity = 96.10%, Specificity = 98.34% for large windows (8 s) Sensitivity = 99.21%, Specificity = 99.68% Detection rate of shockable rhythms (VF and VT) on VFDB: very small windows (2 s) Sensitivity = 94.68%, Specificity = 92.77% for large windows (8 s) Sensitivity = 98.56%, Specificity = 99.08%