Postoperative delirium prediction using machine learning models and preoperative electronic health record data

Figures

• 

  Fig. 1. Inclusion flow diagram

• 

  Fig. 2. Model AUC-ROC curves and calibration plots. A Receiver Operating Characteristic curves for five POD prediction models. B Calibration plots for five POD prediction models. XGBoost (orange), Neural network (blue), Clinician-guided regression (green), ML hybrid regression (red), AWOL-S (purple)

• 

  Fig. 3. Visualization of decisions made by the XGBoost algorithm. A Top 20 most influential variables used by XGBoost. Interpretation: Each dot represents a variable for an individual patient instance. Variables pictured to the right side of the y-axis influenced the model to predict delirium, whereas variables to the left of the y-axis influenced the model against prediction of delirium. Red signifies a higher absolute value (numeric variables) or yes/present (categorical variables), and blue signifies a lower absolute value (numeric variables) or no/absent (categorical variables). For example: Higher age (red color) influenced the model to predict delirium (right of y-axis), whereas lower age (blue color) influenced the model toward prediction of no delirium. Decision path for a true negative (B) and a true positive (C) delirium prediction by XGBoost for two individual patients. Interpretation: The algorithm begins at the center of the x-axis with a baseline value. The model considers each variable along the y-axis one at a time (values shown in parenthesis), to influence the model toward making a positive (vertical line moves toward the right) or negative (vertical line moves toward…

• 

  Fig. 4. Role of automated delirium screen in our institution’s postoperative delirium prevention care pathway. Figure legend: A high-risk delirium screen triggers a set of care modifications in the preoperative, intraoperative, and postoperative phases of care. Particularly in the postoperative phase, these modifications require time and input from busy clinicians