Prolonged Length of Stay After Lung Resection: A Prediction Model That Loses Accuracy Across Hospitals, with No Added Benefit from Blood Pressure Data

Abstract

Background/Objectives: It is uncertain whether intraoperative blood pressure (BP) summaries improve transportable prediction of prolonged postoperative length of stay (PLOS) after lung resection. We developed and temporally validated baseline and BP-enhanced models and evaluated their unchanged performance in the Korean VitalDB cohort. Methods: This retrospective prediction-model study used 536 local patients (249 PLOS events) for development and 184 later patients (74 events) for temporal validation. External validation included 626 procedure-restricted VitalDB patients; paired model comparisons used 624 patients with all four BP features. PLOS was a prespecified operational outcome defined as postoperative stay > 4 days. L2-penalized logistic models were assessed using nested cross-validation, discrimination, calibration, Brier score, and decision-curve analysis. Results: Baseline-model AUROCs were 0.714 (95% confidence interval [CI], 0.667–0.756) internally, 0.719 (0.640–0.788) temporally, and 0.623 (0.579–0.666) externally. Corresponding BP-enhanced AUROCs were 0.721, 0.725, and 0.615. BP features changed temporal AUROC by +0.006 (95% CI, −0.009 to +0.021) and external AUROC by −0.007 (−0.022 to +0.008). Temporal calibration slopes were 1.113 and 1.112, whereas external slopes were 0.570 and 0.534. Sensitivity analyses showed no consistent BP increment. Conclusions: The baseline model retained similar discrimination locally but had limited unchanged external transportability. Four intraoperative BP summaries provided no reproducible incremental predictive value. Neither model is ready for direct cross-center implementation without harmonization, updating, and further validation.