Abstract
<title>Abstract</title> <p>Objective: Acute myocardial infarction complicated by cardiogenic shock (AMI-CS) is an evolving critical-care syndrome in which prognosis changes as tissue hypoperfusion, vasoactive support and multi-organ dysfunction develop. We developed and externally evaluated a shock-anchor-aware informatics workflow for dynamic risk modelling and cross-database transport between MIMIC-IV and eICU-CRD. Materials and Methods: Adult first ICU/unit stays were screened in MIMIC-IV 3.1 and eICU-CRD 2.0. AMI evidence, cardiogenic-shock code/text evidence and an operational shock anchor requiring hemodynamic compromise or vasoactive/inotropic support plus recent hypoperfusion evidence defined the analytic cohort. Patient-level models predicted 30-day in-hospital mortality; hourly dynamic models predicted future 72-hour deterioration or death. Bayesian MAP dynamic models were compared with XGBoost and LASSO/logistic comparators using internal testing, frozen bidirectional external validation, transport recalibration, calibration, decision-curve analysis, SHAP, propensity-score source sensitivity and Wasserstein shift diagnostics. Results: The main analysis included 3,045 MIMIC patients (345 deaths) and 3,738 eICU patients (737 deaths). Dynamic testing included 221,222 MIMIC interval rows with 43,236 events and 270,756 eICU interval rows with 130,398 events. Static frozen XGBoost achieved AUROC 0.814 in MIMIC-to-eICU and 0.838 in eICU-to-MIMIC validation. Dynamic frozen XGBoost achieved AUROC 0.717 in MIMIC-to-eICU and 0.931 in eICU-to-MIMIC validation; corresponding Bayesian MAP AUROCs were 0.611 and 0.793. Brier scores, calibration slopes and decision curves showed clinically meaningful source-direction dependence. Conclusion: AMI-CS prediction is best framed as a dynamic transportability problem rather than a static mortality-classification exercise. The workflow provides a retrospective evidence base for risk-monitoring development, but bedside use requires local recalibration, patient-clustered uncertainty estimation, prospective silent-mode evaluation, clinician-centred interface design and governance review.</p>