ICU患者の28日間死亡率予測のためのマルチモデルフュージョン：サブグループ検証と臨床リスク層別化を伴う包括的な後向きコホート研究

Accurate prediction of 28-day mortality in intensive care unit (ICU) patients represents a critical challenge in modern critical care medicine, with profound implications for clinical decision-making, resource optimization, and patient-centered care. Traditional severity scoring systems and conventional machine learning approaches have consistently demonstrated limitations in discriminative performance, calibration accuracy, and generalizability across heterogeneous patient populations.

This retrospective cohort study employed comprehensive data from 654 consecutively admitted adult ICU patients at a tertiary academic medical center between August 2022 and August 2024.We developed and rigorously validated four distinct prediction methodologies: Logistic Regression (LR), Random Forest (RF), Gradient Boosting (GB), and an innovative Multi-Model Fusion (MMF) framework incorporating probability averaging from constituent models. Our evaluation framework encompassed exhaustive discrimination metrics (AUC-ROC, sensitivity, specificity, F1-score), sophisticated calibration assessment (Brier score, calibration curves, Hosmer-Lemeshow test), detailed subgroup analyses across clinically relevantpatient strata, and multi-dimensional feature importance evaluation.

The Multi-Model Fusion paradigm demonstrated statistically superior performance with an AUC-ROC of 0.862 (95% CI: 0.821-0.903), significantly outperforming all individual models (LR: ΔAUC=0.121, p<0.001; RF: ΔAUC=0.034, p=0.018; GB: ΔAUC=0.027, p=0.032). The fusion model achieved an optimal equilibrium between sensitivity (84.6%) and specificity (81.0%) while maintaining exceptional calibration characteristics (Brier score: 0.140; Hosmer-Lemeshow test: χ²=6.28, p=0.616). Remarkable performance consistency was observed across all patient subgroups (AUC-ROC range: 0.815-0.875), encompassing diverse age strata, disease severity spectra, and intervention requirements. SOFA score, APACHE II score, and Glasgow Coma Scale emerged as consistently robust predictors across all feature importance methodologies. Clinically meaningful risk stratification delineated three distinct mortalitycategories: low-risk (<0.3 probability, 5.6% observed mortality), moderate-risk (0.3-0.7, 31.7%), and high-risk (>0.7, 83.3%).

The Multi-Model Fusion framework establishes a clinically actionable and methodologically sophisticated paradigm for 28-day mortality prediction in ICU patients, effectively addressing fundamental limitations of conventional approaches through enhanced discriminative performance, exemplary calibration accuracy, and consistent generalizability across diverse patient populations. These compelling findings strongly advocate for its integration intoICU clinical decision support ecosystems to advance prognostic precision and guide personalized therapeutic strategies.