Interpretable mortality prediction at VA-ECMO establishment in acute myocardial infarction: A multicenter study from...

Despite the growing use of venoarterial extracorporeal membrane oxygenation (VA-ECMO) as a rescue therapy for acute myocardial infarction (AMI), in-hospital mortality remains high. AMI-specific evidence on prediction timing, calibration, and performance across centers remains limited.

To develop and evaluate interpretable in-hospital mortality prediction models for VA-ECMO-supported AMI patients using multicenter data from the CSECLS registry.

We retrospectively analyzed 1,833 patients from 84 centers between 2016 and 2023. One center (center 188; n = 183) was reserved for held-out evaluation, and the remaining 1,650 patients from 83 centers formed the development cohort. Six candidate models were compared using two clinically distinct information windows. The W1 window included baseline and pre-ECMO variables, and the primary W2 window included baseline, pre-ECMO, and ECMO-establishment variables. Model development and selection used center-aware internal-external cross-validation (IECV) within the development cohort, followed by evaluation in the held-out center 188. The performance of the recommended model was assessed using discrimination, calibration, decision curve analysis, comparison with the partial SAVE score, and SHAP-based interpretability.

CatBoost was recommended for the primary W2 window (pooled IECV AUROC, 0.650). In the held-out center 188, CatBoost achieved an AUROC of 0.834 (95% CI, 0.770-0.891) and a Brier score of 0.182 (95% CI, 0.154-0.209). In the secondary W1 analysis, random forest was selected and achieved an AUROC of 0.855 (95% CI, 0.794-0.912) in the same held-out center. Calibration analysis revealed systematic underestimation of baseline risk and conservative probability estimates, highlighting the need for context-specific recalibration. Decision curve analysis suggested potential clinical net benefit, and the model outperformed the partial SAVE score in the same external cohort. SHAP analysis identified invasive mechanical ventilation, norepinephrine dose, ECMO indication category, arterial pH, mean arterial pressure, and pre-ECMO lactate as the leading predictors.

The CatBoost model for the designated primary W2 window provided an accurate and interpretable tool to support early risk assessment, shared decision-making, and precision management in high-risk cardiogenic shock populations. Local recalibration and further multicenter validation are required before clinical use.