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
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Whether 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.
Methods
Using 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 patient’s 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.
Results
In development the pressure-only model reached AUROC 0.907 vs. 0.884 for MAP alone (ΔAUROC +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, ΔAUROC +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 (ΔAUROC +0.035) but could be externally validated in only 39 patients, because those signals are rarely recorded.
Conclusions
The 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.
