Predicting post-cardiac surgery vasopressor use in a large, multicenter national cardiac surgical database

To develop a model to predict post-cardiac surgery vasopressor administration and describe hospital variation in practice.

Retrospective analysis.

Multi-institutional.

All patients who underwent cardiac surgery with cardiopulmonary bypass between 2012 and 2021.

Observational.

The study cohort was divided into a development set (80%) and a validation set (20%). Univariate logistic regression was used to identify variables associated with postoperative vasopressor administration. The least absolute shrinkage and selection operator was used to develop parsimonious models with variables known preoperatively only (preoperative model), as well as preoperative and immediate postoperative variables (postoperative model). Model discrimination and calibration were performed on both the development and validation sets. The study included 106,348 patients across 33 hospitals. The incidence of postoperative vasopressor administration was 29.3% (n = 31,157). Significant interhospital variability in the rate of the outcome was observed, ranging from 1.20% to 69.4% (median, 22.3%). Fixed effects models with patient and surgical variables were developed for postoperative vasopressor administration, with an area under the receiver operating curve of 0.56 and 0.60 preoperatively and postoperatively, respectively. Accounting for the hospital of admission through mixed effects multilevel modeling improved the area under the receiver operating curve to 0.75 and 0.76 preoperatively and postoperatively, respectively.

Post-cardiac surgery vasopressor administration can only be predicted with poor to fair accuracy based on patient and surgical variables alone. Significant institutional variation in the rate of vasopressor administration exists, seemingly unrelated to measured patient and surgical factors, and predictive ability improves substantially when this is considered.