Table 1.

A summary of automated AF detection in ECG signals. A “+” in the Salient features column indicates a positive point. Conversely, a “−” indicates a negative point.

Authors	Data	Digital Biomarkers	AI	Performance in %			Salient Features
				Acc	Sen	Spe
Wang et al., 2020 [183]	MIT-BIHAFDB	WPD followed by multivariate statistical features	ANN	98.8	98.7	98.9	+Good model performance +Cross-validation used −Digital biomarkers required −Only one database used −No blind fold validation
Cao et al., 2020 [184]	CinCchallenge 2017	Data augmentation	DNN	78.35	-	-	+Improving small datasets −Real measurement data are required −Classification models might exploit weaknesses in the augmentation
Marsili et al., 2019 [185]	MIT-BIH AFDB and measurements	Shannon entropy	Threshold	98.1	99.2	97.3	+Hardware implementation +Fast and energy efficient −Local decision making −Impossible to verify the decision
Yao et al., 2019 [186]	CinC challenge 2017	DWT	Multi-scale CNN	98.18	98.22	98.11	+Good model performance −High computational complexity −One type of digital biomarker
Lui et al., 2018 [187]	MIT-BIH AFDB/NSR/Arrhythmia Database	Normalized fuzzy entropy	Threshold	-	-	-	+Focused study −One type of digital biomarker −No AI
Xia et al., 2018 [188]	MIT-BIH AFIB	STFT, SWT	CNN	98.63	98.79	97.87	+Focused study −One type of digital biomarker −No AI −No validation
Kora et al., 2017 [189]	MIT-BIH AFIB	CS-SCHT	LMNN	99.30	96.97	99.43	+Good model performance −Classical machine learning −No validation
Tripathy et al., 2017 [138]	MIT-BIH AFIB	Sample entropy, VMD	DBN	98.27	98.80	97.77	+Good model performance +Index to combine multiple digital biomarkers −Classical machine learning −No validation −Only one database used
Annavarapu and Padmavathi, 2016 [190]	MIT-BIH Arrhythmia Database	CS-SCHT	LMNN	99.50	99.97	98.70	+Good model performance −Classical machine learning −No validation
Abdul-Kadir et al., 2016 [102]	MIT-BIH NSR/AFDB	Dynamic system	ANN, SVM	95.00			+Cross-validation used −Linear digital biomarkers −Classical machine learning
Yuan et al., 2016 [180]	MIT-BIH NSR/AFDB/LTAFDB	-	Autoencoder DL	98.31	96.56	99.04	+Good model performance +Deep learning −No validation
Asgari et al., 2015 [135]	MIT-BIH AFDB	SWT, Log-energy entropy, peak-to-average power ratio	SVM	97.10	97.00	97.10	+2-fold cross-validation −Classical machine learning
Daqrouq et al., 2014 [191]	MIT-BIH AFIB	WPD	PNN	97.92	-	-	+2-fold cross-validation −Classical machine learning −No outcome directed digital biomarker selection
Martis et al., 2013 [192]	MIT-BIH AFDB/Arrhythmia Database	DWT	NB	99.33	99.32	99.33	+10-fold cross-validation +Multiple arrhythmias +Noise considerations −Classical machine learning
Majia et al., 2013 [193]	MIT-BIH Arrhythmia Database	HOS and EMD	Thresholded	-	96	-	+Classic approach −Single decision border −Only one database −No validation
Rincón et al., 2012 [194]	MIT-BIH AFIB	Statistical measures	Fuzzy classifier	-	98.09	91.66	+P-wave detection −Classical machine learning −Only one database −No validation
Lee et al., 2012 [41]	MIT-BIH NSR/AFDB	RMSSD, sample entropy, Shannon entropy	Threshold	98.44	97.63	99.61	+M-health +Event recorder −Local decision making
Fukunami et al., 1991 [195]	Measurement data	Frequency-domain	Statistical analysis	-	91	76	+Classic manuscript −No benchmark test −No AI
Parvaresh and Ayatollahi, 2011 [196]	MIT-BIH AFIB	Autoregressive model	Statistical classifier	-	96.14	93.20	+Early paper −Machine learning −Linear digital biomarkers