Abstract / Summary
Abstract Background Risk assessment in acute myocardial infarction (AMI) and type 2 diabetes should respect the time at which clinical information becomes available. We evaluated whether mean heart rate during the first 24 hours improves prediction of subsequent in-hospital mortality beyond a compact cross-database clinical model. Methods We performed a retrospective 24-hour landmark study using MIMIC-IV for development and the multicentre eICU Collaborative Research Database for external validation. Adults with AMI and type 2 diabetes who remained in intensive care at 24 hours were eligible. The reference logistic model included age, sex, creatinine, and mean oxygen saturation; the augmented model added mean heart rate. Preprocessing and coefficients were frozen in MIMIC-IV. Internal validity was assessed with 500 bootstrap resamples and external uncertainty with 1000 hospital-cluster bootstrap resamples. Results The landmark cohorts comprised 2193 MIMIC-IV patients (336 deaths; 15.3%) and 287 eICU patients from 65 hospitals (35 deaths; 12.2%). Bootstrap-corrected development AUROC was 0.660 for the reference model and 0.690 for the heart-rate-augmented model. In eICU, AUROC increased from 0.691 (95% CI 0.623–0.798) to 0.742 (95% CI 0.671–0.849); the paired increase was 0.051 (95% CI 0.001-0.100). The augmented model had an AUPRC of 0.284, Brier score of 0.099, calibration intercept of -0.303, and calibration slope of 1.187. Adding systolic blood pressure change did not improve performance in the complete-window sensitivity cohort. On a common eICU subset, APACHE IVa achieved an AUROC of 0.809 versus 0.739 for the augmented model, with an imprecise difference. Conclusions A first-day mean heart rate provided a modest, externally validated improvement in post-landmark mortality risk stratification using routinely recorded variables. The model did not outperform APACHE IVa and requires prospective validation before clinical use.