Dynamic prediction of hospital admission with medical claim data

Figures

• 

  Fig. 1. Illustration of feature extraction. Features are extracted at or during the 3 months, 6 months, 12 months, or more prior to the index date. In our Dynamic Random Survival Forest (DRSF) model, the index date is corresponding to the beginning date of each prediction window

• 

  Fig. 2. Illustration of Random Survival Forest (RSF) model. Features and Samples are selected by random for each tree. Log-rank splitting is used to grow the tree. At the end of each branch, a cumulative hazard function is calculated for the selected individuals. Finally, the ensembled estimated cumulative hazard function is calculated by averaging over all the trees

• 

  Fig. 3. Demonstration of the Dynamic Random Survival Forest (DRSF) Model with sliding windows. The black lines represent different subjects at risk. The red triangles represent the onset of adverse event, e.g. hospital admission. The red box represents the prediction window of interest, while the blue box represents the prediction window for model training and validation purpose corresponding to the red box. At the beginning of the red and blue windows, i.e., the index dates, features are auto-extracted for the model building and prediction (see section “Features and Outcome”). The black and the green boxes are the historical sliding windows trained and predicted following the same concepts as the blue and the red box except they were earlier on the time axis. The green boxes and red box are further combined for the ensemble estimation of the hazard function and survival function at different time points within the red box

• 

  Fig. 4. Comparison of the prediction power among DRSF model with 5 windows, Penalized Cox regression model with most recent window (1 window), and the batch-mode RSF model using the testing window from March 1st, 2015 to August 31st, 2015. The black line is the AUC curve of DRSF with 5 windows, the red line is the AUC curve of Penalized Cox regression with 1 window, and the green line is the AUC curve of batch-mode RSF model. The x-axis represents the number of days since March 1st, 2015. The y-axis represents the Area-Under-the-ROC-Curve (AUC)

• 

  Fig. 5. Variables selected in different testing windows with the index date ranging from August 1st, 2014 to March 1st, 2015 by the increment of a month. Highlighted block indicates a specific variable (variable name annotated along the rightmost y-axis) was selected at a specific window (window index date annotated along the x-axis) by the DRSF model with 3 windows. The leftmost y-axis represents the frequency a variable was chosen among eight available windows. The description of each variable is listed in the Additional file 1: Table S3

• 

  Fig. 6. Dynamic monitoring the risk of hospitalization due to congestive health failure. The red dash line represents the instant risk (ensembled hazard rate) of a randomly selected subject who had an admission event at the 120th days in reality (marked as the red triangle on the horizontal line); the grey solid line represents the instant risk of another randomly selected subject who did not have such an event onset during the prediction window in reality