Figure 3. Test Characteristics for Different Computer Algorithms to Automatically Identify Clinically Relevant Bleeding.

Model names that include -DS (down-sampled) were trained on a note set with an equal number of present and absent events (n = 450) and then tested on the full 660-note test set. Those that include -FS (full sample) were trained on the full training set (n = 990) and then tested on the full 660-note test set. CNN indicates convolutional neural network; ET, extra trees; RB, rules based; SVM, support vector machine.