Development and prospective validation of postoperative pain prediction from preoperative EHR data using attention-based set embeddings

Figures

• 

  Fig. 1. Structure of prediction model. Dimensionality of inputs, outputs, and intermediate representations are indicated in parentheses.

• 

  Fig. 2. Performance metrics for prediction of postoperative pain in the retrospective cohort. Performance was assessed by AUC for NRS > 4 (A) and >6 (B), RMSE (C), and correlation (D) for POPS, along with a table of metrics (E). Shaded areas, error bars, and values represent ±1 standard deviation. Colors indicate different postoperative days. Boxplots (C, D) represent the median (central line), interquartile range (bounds of box) and 1 standard deviation (whiskers). RMSE Root-mean square error, NRS Numeric Rating Scale, AUC area under the curve.

• 

  Fig. 3. Performance metrics for prediction of postoperative pain in the prospective cohort. For POPS and clinician predictions, we present Receiver Operating Curves (ROCs) for NRS > 4 and >6 with Area Under the Curve (AUC) (A–J), RMSE (K), and correlation (L). Shaded areas, error bars, and values represent ±1 standard deviation. Boxplots (K, L) indicate the median (central line), interquartile range (bounds of box) and 1 standard deviation (whiskers). RMSE Root-mean square error, NRS Numeric Rating Scale. Red represents POPS while cyan indicate clinician performance.