The dynamics of arterial pressure itself predict intraoperative hypotension beyond its current value: an interpretable additive model validated in 3,069 external patients under a selection-bias-resistant protocol
- 1Hospital Universitario Puerta del Mar, Cadiz, Spain
Abstract
BackgroundWhether predictors of intraoperative hypotension (IOH) add information beyond the mean arterial pressure (MAP) already displayed on the monitor is contested: selection bias in common evaluation designs inflates apparent performance, and the field has called for comparisons against simple MAP-based references under bias-resistant protocols. Existing predictors also depend on proprietary waveform analysis or pulse-contour monitors, restricting both deployment and external validation.
MethodsUsing 807 non-cardiac surgery patients from the open VitalDB database, we derived an additive gradient boosting model (one split per tree: a learned shape function per variable, no interactions) from three variables computable from an arterial line alone: current MAP, its drift from the patients own 20-minute baseline, and the growth of its rolling variance (critical slowing down). Evaluation used patient-level 5-fold cross-validation under a strict protocol - exclusion of the 65-75 mmHg grey zone and of all samples already hypotensive at prediction time - with MAP alone (same learner class) as comparator. The frozen model was then validated, without any refitting, on an independent cohort from another continent (MOVER, University of California Irvine) following a pre-registered plan sealed before external data access.
ResultsIn development the pressure-only model reached AUROC 0.907 vs. 0.884 for MAP alone ({Delta}AUROC +0.023, 95% CI +0.017 to +0.029) at 5 min, with +0.031 and +0.032 at 10 and 15 min, and good calibration (Brier skill +0.418 vs. prevalence). In external validation on 3 069 patients (442 194 samples, 1-minute charting, event prevalence 5.8%), the advantage not only transferred but was larger than in development: AUROC 0.696 vs. 0.638, {Delta}AUROC +0.058 (95% CI +0.051 to +0.064), meeting both pre-registered gates. Discrimination transferred; calibration did not (external Brier skill -0.014), requiring local recalibration. In the unrestricted scenario, where samples already at threshold are retained, the advantage collapsed (+0.007), reproducing the selection effect this paper documents. A secondary model adding pulse-contour cardiac output and stroke volume variation improved development discrimination further ({Delta}AUROC +0.035) but could be externally validated in only 39 patients, because those signals are rarely recorded.
ConclusionsThe dynamics of arterial pressure itself -- drift from a patient-specific baseline and variance growth -- carry predictive information beyond its current value, in a fully interpretable additive model that requires only an arterial line, no waveform access and no proprietary hardware. The advantage is confirmed in a pre-registered frozen-model external validation of over three thousand patients, and is largest at coarse recording cadence, where instantaneous pressure is least informative.
Related
- Dynamic machine-learning model for early prediction of intraoperative hypotension using 5-minute perioperative time-series data: a multicenter retrospective study
- The Effects of Intraoperative Hypotension on Postoperative Acute Kidney Injury and Myocardial Injury: A Systematic Review and modelling study
- Temporal overlap of intraoperative hypotension and hypocapnia and postoperative outcomes after major open intracranial surgery: a time-weighted dual-hit cohort study
- Cumulative Intraoperative Hypotension Burden as a Predictor of Early Quality of Recovery Following General Anesthesia: A Prospective Observational Study
- Intraoperative hypotension and postoperative acute kidney injury or myocardial injury after liver resection: A protocol for a systematic review and meta-analysis
BibTeX
@misc{oyarzun2026dynamics,
title = {{The dynamics of arterial pressure itself predict intraoperative hypotension beyond its current value: an interpretable additive model validated in 3,069 external patients under a selection-bias-resistant protocol}},
author = {R. Oyarzun and P. Hernandez},
howpublished = {medRxiv},
year = {2026},
month = aug,
doi = {10.64898/2026.08.26.26361468},
url = {https://www.medrxiv.org/content/10.64898/2026.08.26.26361468v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - The dynamics of arterial pressure itself predict intraoperative hypotension beyond its current value: an interpretable additive model validated in 3,069 external patients under a selection-bias-resistant protocol
AU - R. Oyarzun
AU - P. Hernandez
PY - 2026
DA - 2026/08/26
PB - medRxiv
DO - 10.64898/2026.08.26.26361468
UR - https://www.medrxiv.org/content/10.64898/2026.08.26.26361468v1
AB - BackgroundWhether predictors of intraoperative hypotension (IOH) add information beyond the mean arterial pressure (MAP) already displayed on the monitor is contested: selection bias in common evaluation designs inflates apparent performance, and the field has called for comparisons against simple MAP-based references under bias-resistant protocols. Existing predictors also depend on proprietary waveform analysis or pulse-contour monitors, restricting both deployment and external validation. MethodsUsing 807 non-cardiac surgery patients from the open VitalDB database, we derived an additive gradient boosting model (one split per tree: a learned shape function per variable, no interactions) from three variables computable from an arterial line alone: current MAP, its drift from the patients own 20-minute baseline, and the growth of its rolling variance (critical slowing down). Evaluation used patient-level 5-fold cross-validation under a strict protocol - exclusion of the 65-75 mmHg grey zone and of all samples already hypotensive at prediction time - with MAP alone (same learner class) as comparator. The frozen model was then validated, without any refitting, on an independent cohort from another continent (MOVER, University of California Irvine) following a pre-registered plan sealed before external data access. ResultsIn development the pressure-only model reached AUROC 0.907 vs. 0.884 for MAP alone ({Delta}AUROC +0.023, 95% CI +0.017 to +0.029) at 5 min, with +0.031 and +0.032 at 10 and 15 min, and good calibration (Brier skill +0.418 vs. prevalence). In external validation on 3 069 patients (442 194 samples, 1-minute charting, event prevalence 5.8%), the advantage not only transferred but was larger than in development: AUROC 0.696 vs. 0.638, {Delta}AUROC +0.058 (95% CI +0.051 to +0.064), meeting both pre-registered gates. Discrimination transferred; calibration did not (external Brier skill -0.014), requiring local recalibration. In the unrestricted scenario, where samples already at threshold are retained, the advantage collapsed (+0.007), reproducing the selection effect this paper documents. A secondary model adding pulse-contour cardiac output and stroke volume variation improved development discrimination further ({Delta}AUROC +0.035) but could be externally validated in only 39 patients, because those signals are rarely recorded. ConclusionsThe dynamics of arterial pressure itself -- drift from a patient-specific baseline and variance growth -- carry predictive information beyond its current value, in a fully interpretable additive model that requires only an arterial line, no waveform access and no proprietary hardware. The advantage is confirmed in a pre-registered frozen-model external validation of over three thousand patients, and is largest at coarse recording cadence, where instantaneous pressure is least informative.
ER -