Temporal Validation and Simplification of Mortality Prediction for Perioperative Cardiopulmonary Resuscitation

Perioperative cardiac arrest is associated with high mortality, but population-level estimates of outcomes after perioperative cardiopulmonary resuscitation (CPR) do not account for differences in risk among individual patients. We previously developed the CPR OutcoMes Prediction for Arrest in Surgical Settings (COMPASS) model to estimate 30-day mortality after cardiac arrest requiring CPR on the day of surgery. The model incorporates 33 preoperative predictors and has not been evaluated in a temporally distinct cohort. We therefore sought to temporally validate COMPASS and to develop and temporally validate a simplified point-based mortality score.

We performed a retrospective prognostic study of adults undergoing noncardiac surgery who experienced cardiac arrest requiring CPR on postoperative day 0 in the American College of Surgeons National Surgical Quality Improvement Program. The primary outcome was 30-day mortality. The published COMPASS model, developed using 2012-2023 data, was applied without refitting to the 2024 cohort. A simplified point-based score was developed in the 2012-2023 cohort and temporally validated in 2024. Performance was assessed using discrimination, calibration, and Brier score.

The 2024 temporal validation cohort included 429 patients, of whom 274 (63.9%) died within 30 days. The COMPASS model had an area under the receiver operating characteristic curve (AUROC) of 0.82 (95% CI, 0.78-0.86), calibration intercept of 0.24, calibration slope of 1.10, and Brier score of 0.17. The simplified 3-variable score incorporating age, American Society of Anesthesiologists physical status, and case urgency had an AUROC of 0.81 (95% CI, 0.77-0.86), calibration intercept of 0.19, calibration slope of 1.25, and Brier score of 0.17.

In a temporally distinct cohort, the published COMPASS model retained predictive performance for 30-day mortality after perioperative CPR. The simplified score showed similar performance. These findings support further evaluation of both approaches as tools for individualized estimation of perioperative CPR mortality.