Early 6-hour landmark prediction of in-hospital death within 28 days of ICU admission in critically ill patients with acute myocardial infarction: comparative modeling with cross-database external validation

Abstract Background Early prognostic models for critically ill patients with acute myocardial infarction (AMI) often use first-day or full-stay data, and the incremental value of nonlinear machine learning over transparent regression remains uncertain. Methods We reconstructed a retrospective true 6-hour landmark cohort in the Medical Information Mart for Intensive Care IV (MIMIC-IV) for development and internal testing, followed by one correction-specific, one-time external evaluation in the eICU Collaborative Research Database (eICU-CRD) after model locking. Eligible adults had current-admission AMI and were alive and under intensive care unit (ICU) observation at 6 h. The endpoint was index-hospital death after the landmark and within 28 days of ICU admission. We compared a first-available baseline logistic model, a parsimonious logistic model, and compact 0-6-hour logistic and LightGBM models. The compact models used the same 14 predictors. We assessed discrimination, Brier score, calibration, paired differences, and decision curves. Results The MIMIC risk set contained 8,142 admissions and 1,200 events; the untouched internal test set contained 2,436 admissions and 359 events. The eICU risk set contained 10,435 admissions and 799 events. In internal testing, compact logistic regression and LightGBM achieved areas under the receiver operating characteristic curve (AUROCs) of 0.785 (95% confidence interval [CI] 0.761–0.810) and 0.798 (0.774–0.821), areas under the precision-recall curve (AUPRCs) of 0.437 and 0.436, and Brier scores of 0.105 and 0.104, respectively. Their paired AUROC difference was 0.0126 (95% CI -0.0008 to 0.0266; P = 0.0717). In eICU, compact logistic regression and LightGBM achieved AUROCs of 0.847 (0.834–0.860) and 0.851 (0.839–0.864), AUPRCs of 0.362 and 0.357, and Brier scores of 0.059 and 0.059. The external AUROC difference was 0.0041 (95% CI -0.0030 to 0.0118; P = 0.2891), while AUPRC and Brier score did not improve. Conclusions A prespecified strict 6-hour landmark framework was feasible for early cross-database prediction, but it should not be interpreted as an optimal or clinically established prediction time. LightGBM offered limited or uncertain incremental value over same-feature logistic regression. Prospective local validation remains necessary before clinical use.