Dynamic machine learning for time-updated prediction of death within the next 24 hours in critically ill patients with acute myocardial infarction

Critically ill patients with acute myocardial infarction (AMI) admitted to the intensive care unit (ICU) may deteriorate rapidly, yet most existing prognostic tools rely on static admission information and do not update risk as clinical status evolves. We developed and externally validated a dynamic landmark-based machine-learning model to predict in-hospital death within the next 24 hours during ICU stay. Adult ICU stays with AMI were identified in eICU-CRD for model development and internal validation and in MIMIC-IV for external validation. Repeated 6-hour landmarks were generated, and a dynamic LightGBM model was compared with dynamic penalized logistic regression, static first-window LightGBM, dynamic SOFA, an AMI nomogram, APS III/APACHE III, and OASIS. Internal validation included 109,675 eligible landmarks from 10,613 ICU stays; external validation included 52,946 landmarks from 4,505 ICU stays. Dynamic LightGBM achieved the best internal performance (AUROC 0.892, AUPRC 0.231, Brier score 0.0142) and remained the strongest model externally (AUROC 0.868, AUPRC 0.215, Brier score 0.0182). External recalibration improved calibration while preserving discrimination. Decision-curve analysis showed the highest net benefit across clinically relevant thresholds, and the model provided earlier warning with the lowest alert burden (28.4 alerts per 100 patient-days). Performance remained stable across prespecified high-acuity subgroups and sensitivity analyses. These findings support dynamic landmark-based machine learning as a practical framework for short-term mortality monitoring in critically ill patients with AMI.