Table 5:

Attention network performance of DL2 (Class balance) model on the Duke dataset and the Helsinki dataset, as measured by averaged AUC (SD in parentheses). Computation of “Attention AUC” used agreement between each thresholded score and human annotations per channel per epoch; “Attention AUC per epoch” uses agreement of at least one channel score exceeding threshold with human annotation.

Dataset	Attention AUC	Attention AUC per epoch
Duke dataset	0.811 (0.096)	0.927 (0.058)
Helsinki dataset	0.701 (0.107)	0.807 (0.167)