CONGRESS ABSTRACT

Background: sub-phenotyping patients with cardiogenic shock (CS) through a non-traditional clustering method may represent a significant step forward in precision medicine, enhancing clinical outcomes in this heterogeneous and high-mortality condition. We aimed to apply an unsupervised machine learning approach integrating clinical and imaging data (including advanced echocardiography) to identify CS sub-phenotypes associated with different outcomes and treatment responsiveness, beyond etiology. Methods: in this multicenter, observational study we analyzed 172 patients who were prospectively enrolled from 2021 and diagnosed with CS (Society and Cardiovascular Angiography Intervention (SCAI) stage C to E) at admission. The computational analysis comprised an exploratory statistical evaluation followed by a clustering process using the K-Means algorithm. Dimensionality reduction was performed using Principal Component Analysis to visualize clusters in a two-dimensional space. Phenotypes were further stratified according to the SCAI staging system. Results: an optimal cluster number of 5 was determined and the phenotypes identified by the machine learning algorithm were labeled from I to V based on their clinical characteristics at presentation (Table 1). In-hospital mortality rates increased progressively across the phenotypes, with rates of 25%, 32%, 39%, 41% and 60% for Phenotype I to V, respectively (Figure 1). In turn, the 5 phenotypes exhibited different mortality risks at each SCAI stage (Figure 2). The Phenotype I, that showed the lowest mortality, had a higher mean arterial pressure, normal kidney function and a slight increase in left ventricle (LV) diameter with a moderate reduction in ejection fraction (EF) and global longitudinal strain. Phenotype II, also associated with lower mortality, exhibited similar metabolic and hemodynamic features but, interestingly, marked LV dilation and dysfunction and normal right ventricular (RV) function. Phenotype IV and V, associated with the highest mortality, showed elevated lactate and congestion but only a mild reduction in LV EF with normal LV dimension and RV function, suggesting that cardiac function differ between phenotypes. Conclusions: we identified 5 CS phenotypes through a machine learning approach, combining echocardiographic data with clinical variables. These subgroups, characterized by unique mortality risks and features, if further validated, could enhance treatment strategies in CS patients.