Table 2.

Comparison of the pretraining methods depending on the size of the downstream train set.

Pretraining method	25% train	50% train	75% train
None (random weight initialization)	.670 (± .013)	.712 (± .010)	.731 (± .019)
Beat classification	.739 (± .014)	.763 (± .011)	.779 (± .014)
Rhythm classification	.707 (± .018)	.727 (± .028)	.767 (± .012)
Heart rate classification	.722 (± .010)	.749 (± .018)	.766 (± .011)
Future prediction	.694 (± .014)	.734 (± .011)	.758 (± .013)

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. We examine 3 sizes of the train set as a proportion of the entire data set: 25%, 50% and 75% (original split). Pretraining allows models to be trained on less data and still achieve the same degree of performance as the same models that are not pretrained.