Dynamic Landmark-Based Prediction of Sepsis Using Interpretable and Balanced Machine Learning Models in Respiratory-Supported Critically ill Patients

Background Early recognition of sepsis in critically ill patients remains challenging due to dynamic physiological changes and nonspecific clinical presentation. Most prediction models rely on static or continuously updated data streams without explicitly accounting for evolving risk over clinically meaningful time intervals. The study aimed to develop and evaluate a landmark-based dynamic machine learning framework to predict sepsis within a 6-hour horizon among respiratory-supported intensive care unit (ICU) patients. Methods This is a secondary analysis using data from the MIMIC-IV database. Adult intensive care unit patients receiving respiratory support were evaluated at four landmarks (6, 12, 18, and 24 hours). At each point, sepsis-free patients were used to predict sepsis onset within the next 6 hours. Models included logistic regression, random forest, and XGBoost. The patient–level train–test splitting and group cross-validation prevented information leakage. Performance was assessed using discrimination, classification metrics, and calibration. A balanced ensemble approach addressed class imbalance in sensitivity analysis, and interpretability was examined using permutation importance and regression effect estimates. Results A total of 41,871, 39,912, 36,472, and 31,367 patients were included at the 6, 12, 18, and 24-hour landmarks, respectively. Sepsis incidence declined from 1.48% to 0.37% across time points. Model performance varied, with the 18-hour landmark showing the best balance between discrimination and clinically meaningful operating characteristics. Logistic regression achieved the highest discrimination in the primary analysis (AUROC = 0.78), while random forest performed best in sensitivity analyses (AUROC = 0.77). Both consistently identified the 18-hour landmark as optimal, indicating that temporal risk structure outweighed algorithm choice. Calibration was checked overall but showed overestimation at higher predicted risks. Key predictors reflected respiratory, hemodynamic, neurological, and comorbidity factors. Conclusions Landmark-based dynamic modelling provides a clinically interpretable and temporally informed strategy for early sepsis prediction in respiratory-supported intensive care unit patients. The consistent identification of the 18-hour window as the most informative prediction point suggests that intermediate ICU time frames may offer the best balance between timeliness and predictive stability. Further work should focus on recalibration, threshold optimization, and external validation before clinical implementation.