DP27: PREDICTING MORTALITY AT THE MOMENT OF HEMODYNAMIC COLLAPSE: DEVELOPMENT AND INTERNAL VALIDATION OF THE PRESSOR-6 MODEL
Shane W Harrington, MS1; Justin Bogle2; Eelya Sefat3
1FIU Herbert Wertheim College of Medicine; 2University of Florida; 3Campbell University School of Osteopathic Medicine
Introduction/Background: Vasopressor initiation represents a physiologic inflection point signaling progression to shock and sharply increased mortality risk. Existing ICU severity scores, including APACHE IV/IVa, quantify baseline illness burden but are not designed to capture early deterioration following escalation to vasopressor therapy. The first hours after vasopressor initiation represent a clinically actionable window during which trajectory may still be modifiable. We hypothesized that routinely available ICU data within 6 hours of vasopressor initiation could enable accurate, calibrated early mortality prediction through machine learning to support threshold-based triage strategies.
Methods: We performed a retrospective predictive modeling study using the multicenter eICU Collaborative Research Database. Adult ICU patients receiving vasopressors with ≥6 hours of available physiologic data were included. The index time (t0) was defined as first vasopressor administration, and predictions were generated at t0 + 6 hours.
Predictors included demographics, laboratory values, respiratory variables, and vasopressor characteristics summarized within the 0–6 hour window. A gradient-boosted decision tree model (XGBoost) was developed using subject-level splitting to prevent information leakage and evaluated in a held-out internal validation cohort. Performance was compared with logistic regression and APACHE IV/IVa where available.
Discrimination was assessed using AUROC and AUPRC. Calibration was evaluated using Brier score and calibration plots. Prespecified probability thresholds were evaluated in the internal validation cohort to characterize clinically interpretable operating characteristics.
Results: The internal validation cohort included 3,766 ICU stays with 26.2% in-hospital mortality. PRESSOR-6 achieved an AUROC of 0.815 (95% CI 0.801–0.829) and AUPRC of 0.631, outperforming logistic regression (AUROC 0.749) and APACHE IV/IVa (AUROC 0.763). Calibration demonstrated close agreement between predicted and observed mortality (Brier score 0.148).
Threshold-based analysis demonstrated clinically meaningful tradeoffs. A balanced operating threshold achieved sensitivity 0.77 and specificity 0.80. A high-sensitivity threshold approached 0.90 sensitivity, while a top 10% enrichment strategy identified a small subgroup with markedly elevated observed mortality. At comparable operating points, APACHE IV/IVa demonstrated lower discrimination and less favorable threshold behavior.
Discussion/Conclusion: Anchoring prediction to the hemodynamic inflection point of vasopressor initiation enables dynamic mortality risk stratification during an early, modifiable window of shock progression. In internal validation, PRESSOR-6 demonstrated superior discrimination and clinically interpretable threshold performance compared with APACHE IV/IVa. Early post-escalation risk AI modeling or widening the time frame of physiological data collection for modeling may provide a framework for structured triage, escalation, and monitoring strategies in critically ill patients and warrants prospective evaluation.