Table 4.

Comparison of the pretraining methods depending on the architecture of the model (i.e. residual network).

Pretraining method	ResNet-18v2	ResNet-34v2	ResNet-50v2
None (random weight initialization)	.731 (±  .019)	.764 (± .012)	.708 (± .023)
Beat classification	.779 (± .014)	.794 (± .018)	.775 (± .015)
Rhythm classification	.767 (± .012)	.775 (± .020)	.760 (± .008)
Heart rate classification	.766 (± .011)	.771 (± .008)	.761 (± .019)
Future prediction	.758 (± .013)	.761 (± .014)	.743* (± .010)

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. Employing the ResNet-34v2 improves the performance of every pretraining method. We suspect that ResNet-34v2 lies in a sweet spot between model complexity and performance, whereas ResNet-18v2 underfits and ResNet-50v2 overfits to the training data.

*Due to a spike in the model complexity, we only pretrain the first 3 stages of the ResNet-50v2.