Identifying Cardiogenic Shock Sub-Phenotypes with Machine Learning: A Multicenter Study Combining Clinical and Echocardiographic Data

Background: Subphenotyping cardiogenic shock (CS) patients using nontraditional clustering methods represent a step toward precision medicine, potentially improving outcomes in this heterogeneous and high-mortality condition. Objectives: This study aimed to apply an unsupervised machine learning approach to integrate clinical and advanced echocardiographic data, identifying CS subphenotypes associated with different outcomes and features, beyond etiology. Methods: This multicenter observational study prospectively analyzed 172 patients admitted to cardiac intensive care units with overt CS, from 2021. An exploratory statistical analysis preceded patient clustering using the Elbow Method and K-Means algorithm, based on clinical presentation. Dimensionality reduction was performed with principal component analysis. Phenotypes were further stratified according to the Society for Cardiovascular Angiography and Interventions stages. Results: Five distinct phenotypes (I–V) were identified, showing progressively increasing in-hospital mortality rates: 25% (I), 32% (II), 39% (III), 41% (IV), and 60% (V). Kaplan-Meier analysis demonstrated a stepwise increase in mortality risk. Phenotypes IV and V had significantly higher mortality than phenotype I (HR: 2.78 [95% CI: 1.07-7.19] and HR: 2.80 [95% CI: 1.10-7.14]; P < 0.05). Mortality prediction remained independent after adjustment for confounding factors, and independently of Society for Cardiovascular Angiography and Interventions stage. Phenotype I had the lowest mortality, with higher arterial pressure and moderate left ventricular (LV) dysfunction, whereas phenotype II exhibited marked LV failure. Oppositely, phenotypes IV and V had severe congestion despite only mild LV impairment. Conclusions: Machine learning, newly integrating echocardiographic data, identified 5 distinct CS phenotypes, each with unique clinical/echocardiographic features and mortality risks. These insights could support personalized treatment strategies in CS patients, pending further validation.