Figure 2.

Normalization and Data Augmentation Improves the Performance and Stability of the Model

(A and B) The dynamics of training and testing loss during 100 epochs of training process for models applying no normalization and data augmentation (using raw signal) and both normalization and augmentation. Models that achieved the lowest test loss were used as final models to avoid underfitting or overfitting to the training set. The lowest test loss achieved during training is denoted and marked by red crosses. (A) Loss during training without augmentation and normalization. (B) Loss during training with both normalization and augmentation.

(C) Comparison of the performance of the model without both normalization and augmentation (original), with only normalization, with only augmentation, and with both operations.

(D). Pairwise AUROC comparison of the performance by models using original records and with both augmentation and normalization from 1,000,000 bootstrapping operations. The red dashed line denotes a baseline where the performances (AUROCs) of two operations are equal to each other.