Transportability, calibration drift, and recalibration of intraoperative mean arterial pressure trajectory phenotypes for prolonged postoperative length of stay.

Abstract

High-frequency intraoperative monitors support medical-informatics research, but temporal blood-pressure data are often reduced to threshold summaries. To evaluate the transportability of a relative-time mean arterial pressure (MAP) trajectory representation and associated prediction models across MOVER and VitalDB. We derived K = 4 phenotypes from 100-bin MOVER trajectories, projected VitalDB trajectories to frozen MOVER centers, and evaluated nested penalized logistic models. Sensitivity analyses examined K = 2, bootstrap stability, projection distance and ambiguity, duration-frequency matching, cross-fitted recalibration, locally developed VitalDB models, and alternative LOS definitions. The cohorts included 2,298 MOVER patients (571 events) and 1,237 VitalDB patients (371 events). After 15-min duration-stratum frequency matching (716 patients per cohort; median duration 223 vs 220 min), persistent low-MAP showed the same direction of association across cohorts, although the VitalDB estimate was borderline (MOVER adjusted OR 1.90, 95% CI 1.11-3.27; VitalDB adjusted OR 1.76, 95% CI 0.99-3.13; P = 0.056), whereas early-dip remained inconsistent. K = 4 offered negligible external discrimination gain over K = 2. Original external Model 3C had AUROC 0.671 and Brier score 0.242; 10-fold cross-fitted intercept recalibration improved Brier score to 0.193. A locally developed VitalDB Model 2 achieved AUROC 0.783 and Brier score 0.163. Relative-time MAP phenotypes are algorithmically projectable but not uniformly semantically transportable. Persistent low-MAP was the most robust pattern; local model development or recalibration is preferable for absolute-risk use.