Objective: To explore the feasibility of real-time mortality risk assessment for ICU patients.
Design/methods: This study used retrospective analysis of mixed medical/surgical intensive care patients in a university hospital. Logistic regression was applied to 7048 development patients with several hundred candidate variables. Final models were selected by backward elimination on top cross-validated variables and validated on 3018 separate patients.
Results: The real-time model demonstrated strong discrimination ability (Day 3 AUC=0.878). All models had circumstances where calibration was poor (Hosmer-Lemeshow goodness of fit test p < 0.1). The final models included variables known to be associated with mortality, but also more computationally intensive variables absent in other severity scores.
Conclusion: Real-time mortality prediction offers similar discrimination ability to daily models. Moreover, the discrimination of our real-time model performed favorably to a customized SAPS II (Day 3 AUC=0.878 vs AUC=0.849, p < 0.05) but generally had worse calibration.