Abstract
Introduction: Preoperative registry variables are widely used for bariatric risk counseling; however, the ceiling of patient-level prediction using routinely collected data remains uncertain.
Objective: To benchmark the performance of logistic regression versus gradient-boosted decision trees for prediction of 30-day outcomes after bariatric surgery using contemporary MBSAQIP data and to evaluate achievable discrimination, calibration, and clinical utility when models are restricted to standard preoperative registry variables.
Methods: We analyzed adults undergoing bariatric surgery in the 2020–2023 MBSAQIP Participant Use Files. Models were developed using 2020–2022 data and temporally validated in an independent 2023 cohort. Using 30 fixed preoperative predictors, we assessed discrimination (AUROC), calibration (Brier score; calibration intercept and slope), and decision-curve utility across 1–10% risk thresholds. Performance was examined by procedure type and procedural intent.
Results: Among 702,614 patients, several outcomes were rare (leak 0.2%, mortality 0.03%). Discrimination was modest across outcomes and subgroups (typical AUROC 0.58–0.62), and XGBoost did not consistently outperform logistic regression. For sleeve gastrectomy readmission, both models demonstrated acceptable population-level calibration (Brier 0.022; slopes near 1), but net clinical benefit was limited. For rare outcomes such as leak, calibration was poor, and no meaningful net benefit was observed.
Conclusion: Using standard preoperative MBSAQIP variables, both logistic regression and XGBoost provide limited individual-level prediction of 30-day outcomes, particularly rare events. Clinically actionable improvement will likely require richer perioperative inputs rather than increased algorithmic complexity alone.