Performance analysis for best-performing LSTM models. The area under the receiver operating characteristic curve (ROC-AUC) demonstrated a score of 0.98 for both the hematology-oncology model (B) and the multispecialty model (C) and a score of 0.95 for the cardiothoracic model (A). The AUC-PR (D-F) is represented for each model type. The ideal threshold for the cardiothoracic model (D) was identified as 0.39 (0.38 for the test set), compared with 0.47 (0.42 for the test set) for the multispecialty models (F) and 0.41 (also 0.37 for the test set) for the hematology-oncology model (E). Confusion matrices for all the models that have been trained are shown (G-I). Note that the “true negative” class (referring to those patients who were accurately identified as not having received any PCs) was disproportionately represented in all data sets, and the multispecialty model included a disproportionately large sample size.