Background

Acute myocardial infarction complicated by cardiogenic shock (AMI-CS) carries a high risk of in-hospital death. Randomized evidence does not support routine VA-ECMO use in unselected AMI-CS populations. Objective and interpretable risk stratification is clinically important when advanced mechanical circulatory support is considered.

Methods

This was a retrospective, multi-database model development and validation study with exploratory application in a local VA-ECMO-treated AMI-CS cohort. An in-hospital mortality prediction model was developed in AMI-CS patients who did not receive ECMO treatment using the MIMIC-IV database. Candidate measurements were summarized using the worst values accumulated during the first 24 h after ICU admission. Variable shrinkage was performed using LASSO logistic regression with 10-fold cross-validation, and the model was established using multivariable logistic regression. Internal validation included the original fixed-structure bootstrap and an additional full-pipeline bootstrap sensitivity analysis. The latter repeated preprocessing, LASSO selection, post-LASSO handling, and model refitting within each resample. External validation was performed in the eICU cohort. The model was then applied exploratorily to the local VA-ECMO-treated AMI-CS cohort to describe model transportability and the relationship between observed and predicted outcomes.

Results

The MIMIC-IV development cohort included 355 patients, of whom 127 died in hospital (35.77%). LASSO selected 8 clinical predictors. The apparent AUC was 0.829 (95% CI 0.785–0.874). After the final model was refitted in 200 fixed-structure bootstrap resamples, the optimism-corrected AUC was 0.814. In 500 full-pipeline bootstrap resamples, all repetitions were valid, the mean optimism was 0.0436, and the optimism-corrected AUC was 0.786 (empirical 2.5th to 97.5th percentile range: 0.745–0.832). The AUCs from the MIMIC-IV complete-case and multiple-imputation sensitivity analyses were 0.820 and 0.840, respectively. In the eICU complete-case cohort, the AUC was 0.806 (95% CI 0.691–0.921), O/E was 0.733, the calibration intercept was − 0.986, and the calibration slope was 0.671. After multiple imputation in all 224 eligible eICU patients, the AUC was 0.771, O/E was 0.754, the calibration intercept was − 0.756, and the calibration slope was 0.528. The local cohort included 40 VA-ECMO-treated patients. Overall observed mortality was 42.5% (95% CI 27.0-59.1), mean predicted mortality was 59.2% (95% CI 50.6–67.2), and O/E was 0.717 (95% CI 0.418–1.149).

Conclusions

This study developed and externally validated a mortality risk model for patients with acute myocardial infarction complicated by cardiogenic shock. The model may provide a research reference for early risk stratification and individualized assessment. Uncertainty in variable selection and cross-population calibration drift limit its direct clinical application. Prospective multicenter studies are required for further validation and to evaluate its role in VA-ECMO patient selection.