Table 3.

Results comparing the prediction performance between the proposed model and state-of-the-art models.

Author	Year	Group	Database	Features	Classifier	Explainable	Before CAa	Performance
Churpek et al [14]	2016	Non-CA: 253,547; CA: 424	Clinical database	Time since ward admission, demographics, hospitalization history, vital signs, and laboratory results	RFb	Yes	0 h, current point	AUROCc=0.83
Kwon et al [18]	2018	Non-CA: 45,539; CA: 396	Clinical database	Vital signs	RNNd	No	0 h, current point	AUROC=0.85; AUPRCe=0.04
Layeghian Javan et al [16]	2019	Non-CA: 2681; CA: 79	MIMICf-III [17]	Time interval and statistical features using vital signs and clinical latent features	Stacking	No	1 h	AUROC=0.82
Proposed method	N/Ag	Non-CA: 1899; CA: 82	MIMIC-IV [21]	Cosine similarity and statistical features using vital signs and clinical latent features	LGBh	Yes	1 h	AUROC=0.86 AUPRC=0.58

^aCA: cardiac arrest.

^bRF: random forest.

^cAUROC: area under the receiver operating characteristic curve.

^dRNN: recurrent neural network.

^eAUPRC: area under the precision-recall curve.

^fMIMIC: Medical Information Mart for Intensive Care.

^gN/A: not applicable.

^hLGB: gradient boosting ensemble of decision trees.