Abstract
<title>Abstract</title> <p> <bold>Background.</bold> Active fluid removal (de-resuscitation) is frequently required after the resuscitation of critically ill patients, but it can precipitate haemodynamic instability, and no validated tool estimates this risk at the moment fluid removal is initiated. We developed and externally validated a model to predict haemodynamic intolerance within 24 hours of the first intravenous furosemide dose in haemodynamically stable intensive care patients, and examined whether its discrimination and calibration transport across time and across populations. <bold>Methods.</bold> In a retrospective cohort study reported in accordance with the TRIPOD + AI statement, the decision point (t0) was the first intravenous furosemide administration in adults with no vasopressor in the preceding 6 hours. The primary outcome was haemodynamic intolerance within 24 hours, defined as new vasopressor initiation or a mean arterial pressure below 60 mmHg recorded on at least three occasions. Twenty routinely collected predictors available in both databases were used. Penalised (ridge) logistic regression and gradient-boosted trees were developed in MIMIC-IV and evaluated by internal cross-validation, internal temporal validation, and external geographic validation in the eICU-CRD without refitting. Discrimination, calibration, and decision-curve net benefit were assessed, and miscalibration was addressed by out-of-sample logistic recalibration. <bold>Results.</bold> The development cohort comprised 14,294 admissions (27.9% with intolerance) and the external cohort 21,677 (31.4%). Discrimination was moderate and stable across all settings (gradient-boosting area under the receiver-operating-characteristic curve 0.764 [95% CI 0.755–0.773] for internal cross-validation, 0.733 [0.710–0.754] for temporal validation, and 0.767 [0.760–0.774] for external validation). In contrast, calibration shifted predictably with event prevalence — over-prediction when prevalence fell across the temporal split and under-prediction when it rose in the external cohort — while the calibration slope remained close to, or slightly below, one; a single logistic recalibration largely corrected the shift in both directions. Decision-curve analysis showed net benefit across a clinically plausible range of thresholds. The gradient-boosting and ridge models performed similarly. <bold>Conclusions.</bold> A model predicting haemodynamic intolerance at the de-resuscitation decision showed portable discrimination but prevalence-dependent calibration that was largely correctable by simple recalibration. The model is positioned as decision support and should be recalibrated locally before use; prospective impact evaluation is the essential next step. </p>