Figure 3.

Topic count selection. Average discrimination accuracy (ROC AUC) when predicting additional clinical orders occurring within t followup verification time of the invocation of a pre-authored order set during the first 24 hours of hospitalization for 4820 validation patients. Predictions based on Latent Dirichlet allocation (LDA) topic models trained on 10 655 separate training patients. The standard LDA algorithm requires external specification of a topic count parameter to indicate the number of latent dimensions by which to organize the source data, which varies along this X-axis from 2 to 2048. Peak performance occurs around a choice of 32 topics, degrading once attempting to model > 64 topics.