Machine Learning-Augmented Triage for Sepsis: Real-Time ICU Mortality Prediction Using SHAP-Explained Meta-Ensemble Models

Abstract

Background/Objectives: Optimization algorithms are acknowledged to be critical in various fields and dynamical systems since they provide facilitation in identifying and retrieving the most possible solutions concerning complex problems besides improving efficiency, cutting down on costs, and boosting performance. Metaheuristic optimization algorithms, on the other hand, are inspired by natural phenomena, providing significant benefits related to the applicable solutions for complex optimization problems. Considering that complex optimization problems emerge across various disciplines, their successful applications are possible to be observed in tasks of classification and feature selection tasks, including diagnostic processes of certain health problems based on bio-inspiration. Sepsis continues to pose a significant threat to patient survival, particularly among individuals admitted to intensive care units from emergency departments. Traditional scoring systems, including qSOFA, SIRS, and NEWS, often fall short of delivering the precision necessary for timely and effective clinical decision-making. Methods: In this study, we introduce a novel, interpretable machine learning framework designed to predict in-hospital mortality in sepsis patients upon intensive care unit admission. Utilizing a retrospective dataset from a tertiary university hospital encompassing patient records from January 2019 to June 2024, we extracted comprehensive clinical and laboratory features. To address class imbalance and missing data, we employed the Synthetic Minority Oversampling Technique and systematic imputation methods, respectively. Our hybrid modeling approach integrates ensemble-based ML algorithms with deep learning architectures, optimized through the Red Piranha Optimization algorithm for feature selection and hyperparameter tuning. The proposed model was validated through internal cross-validation and external testing on the MIMIC-III dataset as well. Results: The proposed model demonstrates superior predictive performance over conventional scoring systems, achieving an area under the receiver operating characteristic curve of 0.96, a Brier score of 0.118, and a recall of 81. Conclusions: These results underscore the potential of AI-driven tools to enhance clinical decision-making processes in sepsis management, enabling early interventions and potentially reducing mortality rates.

Keywords: clinical critical decision support; deep learning; emergency department; in-hospital mortality; intensive care unit; machine learning; predictive modeling; sepsis; stacked ensemble model.

Figure 1

Distribution of standardized clinical features.

Figure 2

Overview of the proposed end-to-end clinical AI pipeline for in-hospital mortality prediction for sepsis patients.

Figure 3

Interpretability analysis of the Stacked Ensemble ML model.

Figure 4

Confusion matrix of the prediction accuracy of the proposed Stacked Ensemble model on the internal validation dataset.

Figure 5

SHAP summary highlighting the top 10 clinical features based on mean absolute SHAP values. The CRP/albumin ratio, qSOFA score, and NLR exhibited the highest predictive contribution to the Stacked Ensemble model’s output, reinforcing their pathophysiological relevance in sepsis triage.

Figure 6

ROC curve illustrating the discriminative performance of the proposed Stacked Ensemble model. The dotted line represents the performance of a random classifier, serving as a baseline to show that the model curve.

Figure 7

Heatmap illustrating the mean SHAP values of the top 10 clinical features across TPs, FPs, and FNs in the internal validation dataset.

Figure 8

Enhanced heatmap showing class-wise SHAP feature distribution among correctly and incorrectly classified sepsis cases.