Acute coronary syndrome (ACS) remains a major cause of mortality worldwide. However, the accuracy of current prediction tools for in-hospital cardiac events after an ACS remains insufficient for individualized patient management strategies.

To assess in patients with ACS the feasibility and accuracy of machine learning (ML)-based model using all data available at admission to predict in-hospital cardiac events.

We conducted a sub-study of ADDICT-ICCU registry, an observational prospective study including all consecutive patients admitted to intensive cardiac care unit (ICCU) in 39 centres throughout France between 7 and 22 April 2021. We evaluated 16 clinical, 4 biological and 6 Transthoracic echocardiogram (TTE) features. ML involved automated feature selection with XGBoost then model building by random forest (RF), and hyperparameter tuning was done by repeated cross-validation. The primary outcome was the occurrence of composite outcomes defined by death, resuscitated cardiac arrest or cardiogenic shock requiring medical and/or mechanical haemodynamic support.

Of 1,499 consecutive patients, 765 (mean age 63±15 years, 70% male) were admitted for ACS. The overall in-hospital cardiac events rate for ACS patients was 4.0%. Feature selection was performed using XGBoost with the log-rank–based variable importance, and 4 of the available features at admission were selected for the RF model (4 from TTE) including cardiac output, filling pressures, tricuspid annular plane systolic excursion and pulmonary arterial systolic pressure (Figure 1, Panel A). The ML model exhibited a higher area under the curve compared with TIMI score, GRACE score, and traditional stepwise model score for in hospital major adverse events prediction (ML score: 0.96 vs TIMI: 0.54, GRACE: 0.68, traditional stepwise score: 0.87; all P<0.001) (Figure 1, Panel B).

The ML-model exhibited a higher prognostic value to predict in-hospital cardiac events compared with all traditional scores.