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

Area of Science:

• Cardiology
• Medical Informatics
• Data Science

Background:

• Cardiogenic shock (CS) is a high-mortality condition with significant heterogeneity.
• Precision medicine requires effective patient subphenotyping.
• Nontraditional clustering methods offer a path toward improved CS patient stratification.

Purpose of the Study:

• To apply unsupervised machine learning to integrate clinical and echocardiographic data for CS subphenotyping.
• To identify CS phenotypes associated with distinct clinical features and outcomes.
• To move beyond etiological classification for a more data-driven approach.

Main Methods:

• Prospective observational multicenter study of 172 CS patients.
• Unsupervised clustering using Elbow Method and K-Means algorithm on clinical data.
• Principal Component Analysis for dimensionality reduction and integration of echocardiographic data.
• Stratification based on Society for Cardiovascular Angiography and Interventions (SCAI) stages.

Main Results:

• Five distinct CS phenotypes (I-V) were identified.
• In-hospital mortality progressively increased across phenotypes (25% to 60%).
• Phenotypes IV and V exhibited significantly higher mortality risk compared to Phenotype I, independent of SCAI stage.
• Distinct clinical and echocardiographic profiles characterized each phenotype, including left ventricular (LV) dysfunction and congestion levels.

Conclusions:

• Machine learning successfully identified five novel CS phenotypes integrating clinical and echocardiographic data.
• Each phenotype possesses unique characteristics and differential mortality risks.
• These findings support the development of personalized treatment strategies for CS patients.
• Further validation is necessary to confirm the clinical utility of these data-driven phenotypes.