Dynamic prognostic prediction in sepsis using longitudinal blood gas trajectories: development and external validation

Objective: To develop and externally validate a dynamic prognostic framework for sepsis based on longitudinal arterial blood gas trajectories, and to assess whether ABG-derived dynamic features provide incremental prognostic information beyond conventional clinical variables. Methods: ratio, pH, base excess, lactate, and glucose, summarized in 12-h windows. Latent class growth analysis and growth mixture modeling identified longitudinal ABG trajectories in MIMIC-IV, and the locked models were directly applied to eICU-CRD. Dynamic prediction models were developed using data accrued up to days 3, 5, and 7 with elastic-net logistic regression. Performance was evaluated by ROC-AUC, PR-AUC, Brier score, calibration, and decision curve analysis. Results: After applying the longitudinal ABG completeness criteria, 1,378 patients from MIMIC-IV and 647 from the eICU-CRD external validation cohort were included. Clinically interpretable trajectories were identified for all eight ABG variables, suggesting dynamic physiological heterogeneity in sepsis. For post-landmark in-hospital mortality prediction, models integrating clinical variables, ABG trajectory posterior probabilities, and ABG summary statistics showed moderate discrimination. In MIMIC-IV internal validation, ROC-AUCs using data accrued up to days 3, 5, and 7 were 0.792, 0.796, and 0.808, respectively; the corresponding values in eICU-CRD external validation were 0.658, 0.679, and 0.703. Compared with the clinical reference model, adding ABG-derived dynamic features yielded small-to-moderate improvements in external validation, with ROC-AUC increases of 0.036, 0.044, and 0.055 across the three landmarks. However, gains were smaller in internal validation and not uniformly robust across settings. Incident septic shock analyses were limited by few post-landmark events, particularly in eICU-CRD, and were considered exploratory. Conclusion: Early longitudinal ABG trajectories in sepsis may capture dynamic physiological heterogeneity and provide limited complementary information beyond conventional clinical variables. Given the modest incremental benefit and inconsistent performance across cohorts, ABG-derived dynamic features should be regarded as adjunctive to clinical severity scores and clinical judgment. Their clinical utility requires prospective validation.