Machine Learning Prediction of Early ICU-to-Ward Transfer in Patients with Sepsis: Development, External Validation, and Retrospective Cohort Study

Early identification of patients likely to undergo ICU-to-ward transfer may inform transfer planning in sepsis. We developed LightGBM models using first 24 h ICU data from MIMIC-IV and evaluated them in a temporal internal test set and eICU-CRD. In MIMIC-IV, eligible patients met Sepsis-3 criteria by 24 h and remained alive in the ICU at that landmark. The primary outcome was ward transfer between 24 and 72 h without recorded ICU readmission or death within seven days after transfer. The severity-score-free 55-feature model was the primary model for cross-database evaluation; the original 56-feature model was retained for comparison, with eICU evaluation requiring substitution of APACHE-IVa for APS III. The primary external cohort comprised 13,384 patients from 200 hospitals with an APACHE admission diagnosis of sepsis or septic shock. The primary model achieved internal and external AUCs of 0.800 and 0.745, respectively, versus 0.809 and 0.726 for the full model. External calibration slopes were 0.907 and 0.741, and sensitivities at development-derived thresholds were 0.410 and 0.395, respectively. Low external sensitivity and variation across hospitals limit clinical application. Prospective validation is required before use.