Operative-Time Gradients Reveal Miscalibration in Microsurgical Risk Prediction
TL;DR
Generalized ACS NSQIP morbidity prediction was accurate in aggregate but systematically miscalibrated across the operative-time gradient, overestimating risk in shorter operations and underestimating it in longer operations.
Abstract
Abstract Background General surgical risk models support perioperative counselling, but acceptable overall calibration may conceal clinically important error. We evaluated generalized American College of Surgeons National Surgical Quality Improvement Program (ACS NSQIP) predictions after microsurgical reconstruction and tested whether morbidity calibration varied across operative time. Methods Adult ACS NSQIP cases from 2014 to 2023 were analyzed. Plastic-surgery-service cases meeting exact principal Current Procedural Terminology (CPT) or principal-procedure-text criteria formed the primary cohort. Performance was assessed using observed-to-expected (O:E) ratios, discrimination, and calibration measures. Operative-time calibration was examined by quartiles, deciles, and continuous spline models. Sensitivity analyses used exact principal CPT codes alone and excluded cases with a known morbidity component or reoperation on postoperative day 0 or 1. Results Among 20,604 microsurgery cases, observed and predicted morbidity were similar overall (12.52 vs. 12.79%; O:E, 0.98; 95% confidence interval [CI], 0.94–1.02), although discrimination was modest (area under the receiver operating characteristic curve, 0.639; 95% CI, 0.627–0.651). Mortality was uncommon (21 deaths; O:E, 0.81; 95% CI, 0.50–1.24). Aggregate calibration concealed a graded reversal across operative time. Morbidity was overpredicted in the shortest quartile (8.66% observed vs. 12.87% predicted; O:E, 0.67; 95% CI, 0.61–0.74) and underpredicted in the longest quartile (16.94 vs. 12.96%; O:E, 1.31; 95% CI, 1.22–1.40). The prespecified linear interaction was nonsignificant ( p = 0.183), whereas flexible continuous analyses demonstrated calibration variation before and after global recalibration (both p < 0.001). Findings persisted in both sensitivity analyses. Conclusion Generalized ACS NSQIP morbidity prediction was accurate in aggregate but systematically miscalibrated across the operative-time gradient, overestimating risk in shorter operations and underestimating it in longer operations. Realized operative time should be interpreted as a postoperative marker of incompletely captured procedural complexity and intraoperative course, not as a causal exposure or preoperative predictor.