JAMA Netw Open. 2025 Oct 1;8(10):e2539767. doi: 10.1001/jamanetworkopen.2025.39767.
ABSTRACT
IMPORTANCE: Perioperative cardiac arrest and cardiopulmonary resuscitation (CPR) are associated with significant morbidity and mortality. Despite a growing focus on goal-concordant surgical care and longstanding emphasis on preoperative code status discussions, clinicians lack tools to individualize risk estimates and inform shared decision-making (SDM) regarding perioperative CPR.
OBJECTIVE: To generate and internally validate predictive models for 30-day mortality and nonhome discharge using routinely available preoperative data.
DESIGN, SETTING, AND PARTICIPANTS: A prospective, multicenter, prognostic study of patients within the American College of Surgeons (ACS)-National Surgical Quality Improvement Program (NSQIP), including nearly 700 participating hospitals in the US, from January 1, 2012, through December 31, 2023. Follow-up duration was 30 days. Seven machine learning models were developed using 10-fold cross validation. Participants were patients aged 18 years or older undergoing noncardiac surgery who underwent CPR on the day of surgery.
EXPOSURES: Thirty-three preoperative sociodemographic, clinical, laboratory, and procedural variables were evaluated for their association with 30-day mortality and nonhome discharge.
MAIN OUTCOMES AND MEASURES: The primary outcome was 30-day mortality following CPR. The secondary outcome was nonhome discharge among survivors admitted from home. Performance was evaluated using area under the receiver operating characteristic curve (AUROC), accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and calibration (Brier score and calibration curves). Clinical utility was evaluated using Shapley additive values (SHAP) decision curve analysis (DCA).
RESULTS: Among 6405 patients (median [IQR] age 69 [60-78], 3572 [55.8%] men, 860 [13.4%] Black, 4343 [67.8%] White, and 261 [4.1%] another racial category, including American Indian or Alaska Native, Asian, and Native Hawaiian or Other Pacific Islander), 3710 (57.9%) died within 30 days. The extreme gradient boosting model (CPR Outcome Prediction for Arrest in Surgical Settings [COMPASS]) achieved the best performance (AUROC for mortality = 0.80; 95 % CI, 0.78-0.82; accuracy = 0.73; 95% CI, 0.71-0.75; sensitivity = 0.77; 95% CI, 0.74-0.79; specificity = 0.68; 95% CI, 0.65-0.71; Brier score = 0.18). Among 2478 survivors admitted from home, 822 (33.2%) were discharged to a facility. For nonhome discharge, extreme gradient boosting demonstrated an AUROC of 0.78 (95 % CI, 0.74-0.82), accuracy of 0.76; and Brier score of 0.17. SHAP analysis identified American Society of Anesthesiologists status, case urgency, and frailty as key predictors. DCA indicated greater net benefit of extreme gradient boosting over default strategies (ie, treat all or treat none) across wide threshold ranges.
CONCLUSIONS AND RELEVANCE: In this prospective prognostic study of outcomes following perioperative CPR, extreme gradient boosting generated individualized predictions of outcomes following perioperative CPR that may inform prevention strategies and goal-concordant surgical care.
PMID:41148140 | DOI:10.1001/jamanetworkopen.2025.39767
