Abstract Background Prognosis models based on stepwise regression methods show modest performance in patients with cardiogenic shock (CS). Automated variable selection allows data-driven risk evaluation by recognizing distinct patterns in data. We sought to evaluate an automated variable selection method (least absolute shrinkage and selection operator, LASSO) for predicting 30-day mortality in patients with acute myocardial infarction and CS (AMICS) receiving acute percutaneous coronary intervention (PCI) compared to two established scores. Methods and results Consecutive patients with AMICS receiving acute PCI at one of two tertiary heart centres in Denmark 2010–2017. Patients were divided according to treatment with mechanical circulatory support (MCS); PCI–MCS cohort (n = 220) versus PCI cohort (n = 1180). The latter was divided into a development (2010–2014) and a temporal validation cohort (2015–2017). Cohort-specific LASSO models were based on data obtained before PCI. LASSO models outperformed IABP-SHOCK II and CardShock risk scores in discriminative ability for 30-day mortality in the PCI validation [receiver operating characteristics area under the curve (ROC AUC) 0.80 (95% CI 0.76–0.84) vs 0.73 (95% CI 0.69–0.77) and 0.70 (95% CI 0.65–0.75), respectively, P < 0.01 for both] and PCI–MCS development cohort [ROC AUC 0.77 (95% CI 0.70–0.83) vs 0.64 (95% CI 0.57–0.71) and 0.64 (95% CI 0.57–0.71), respectively, P < 0.01 for both]. Variable influence differed depending on MCS, with age being the most influential factor in the LASSO–PCI model, whereas haematocrit and estimated glomerular filtration rate were the highest-ranking factors in the LASSO–PCI–MCS model. Conclusion Data-driven prognosis models outperformed established risk scores in patients with AMICS receiving acute PCI and exhibited good discriminative abilities. Observations indicate a potential use of machinelearning to facilitate individualized patient care and targeted interventions in the future.