Fig. 5. Generalization and transfer learning results across institutions.

a We train models using both Stanford and TUH datasets, and evaluate on the test set from each institution. We show median ROC curves obtained from 5 runs of each model. X–Y stands for trained on X and evaluated on Y. b We fine-tune our model pre-trained on the large weakly labeled Stanford dataset using the TUH dataset. We compare the median ROC curves with and without pre-training.