Аннотация
Early identification of patients at high risk of in-hospital mortality is a persistent challenge in clinical practice, particularly during infectious disease outbreaks such as COVID-19, when decisions must be made at the time of hospital admission using incomplete information. Although machine learning methods have been widely applied to mortality prediction, many existing models are developed and assessed under conditions that limit their usefulness for early triage, including insufficient attention to probability reliability, temporal generalization, and decisionoriented assessment. This work examines early-stage mortality risk prediction using routinely available clinical data collected at the time of hospital admission. An ensemble of gradient boosting models is developed and analyzed under strict temporal separation to reflect real-world deployment conditions. Particular emphasis is placed on the reliability of predicted probabilities, with post-hoc isotonic calibration applied to improve alignment between predicted risks and observed outcomes. Model performance is assessed using complementary discrimination and calibration metrics, while clinical relevance is examined through decision curve analysis across plausible operating thresholds. Evaluation on an independent, temporally held-out test set shows that ensemble aggregation improves the stability of risk estimates, while calibration yields more consistent threshold-based behavior without altering ranking performance. Decision curve analysis indicates that calibrated predictions provide higher net benefit than default decision strategies across a broad range of early triage thresholds. These findings highlight that, in early clinical decision-making, probability reliability plays a critical role alongside discrimination. The presented framework offers a methodologically robust approach to early in-hospital mortality risk assessment under realistic clinical and temporal constraints.