Table 1.

Comparison of different configurations of the pretraining methods.

Pretraining method				Frame	$F_1$	$F_{1n}$	$F_{1a}$	$F_{1o}$	$F_{1p}$
None (random weight initialization)					.731 (± .019)	.898 (± .005)	.711 (± .027)	.701 (± .017)	.613 (± .062)
Beat classification				512	.769 (± .011)	.911 (± .010)	.760 (± .018)	.758 (± .016)	.647 (± .022)
				2048	.779 (± .014)	.915 (± .007)	.777 (± .014)	.763 (± .014)	.661 (± .040)
				4096	.768 (± .010)	.908 (± .009)	.764 (± .021)	.754 (± .015)	.646 (± .025)
Rhythm classification				512	.742 (± .017)	.896 (± .007)	.721 (± .026)	.716 (± .032)	.636 (± .045)
				2048	.767 (± .012)	.908 (± .004)	.753 (± .020)	.745 (± .018)	.660 (± .026)
				4096	.755 (± .005)	.903 (± .008)	.745 (± .022)	.735 (± .012)	.635 (± .017)
Heart rate classification				512	.766 (± .011)	.915 (± .004)	.759 (± .019)	.756 (± .015)	.635 (± .029)
				2048	.753 (± .013)	.910 (± .005)	.743 (± .037)	.738 (± .011)	.619 (± .039)
				4096	.751 (± .010)	.909 (± .006)	.744 (± .019)	.739 (± .016)	.611 (± .025)

	Context	ns	Offset	Frame
Future prediction	8	4	2	512	.756 (± .008)	.903 (± .007)	.742 (± .011)	.730 (± .017)	.649 (± .021)
	16	8	2	512	.744 (± .016)	.905 (± .005)	.730 (± .027)	.730 (± .009)	.612 (± .041)
	16	8	8	512	.758 (± .013)	.908 (± .005)	.753 (± .021)	.745 (± .012)	.627 (± .026)
	16	16	8	512	.745 (± .013)	.897 (± .006)	.724 (± .024)	.722 (± .009)	.639 (± .034)

For each method, we report the average macro $F_1$ score (and the standard deviation) on our test set for the PhysioNet/CinC Challenge 2017^7,8. Additionally, we report the average $F_1$ score for each class: normal ($F_{1n}$), AF ($F_{1a}$), other ($F_{1o}$) and noisy ($F_{1p}$). Frame refers to the length of an ECG frame, context to the number of frames in the context, ns to the number of negative samples and offset to the distance between the context and the future frame measured in frames. All pretraining methods outperform random weight initialization in predicting every class.