# Phenotyping cardiogenic shock: an insight from the gulf cardiogenic shock registry

## TL;DR

This study uses machine learning to identify four distinct types of cardiogenic shock, each with different risk levels and outcomes, which could help in developing personalized treatments.

## Contribution

The study introduces a novel data-driven framework for risk stratification in cardiogenic shock using unsupervised machine learning.

## Key findings

- Four distinct CS phenotypes were identified with varying mortality rates and clinical characteristics.
- Phenotype 3 had the highest mortality at 78.4%, marked by multi-organ failure.
- The identified phenotypes showed a steep mortality gradient and distinct SCAI shock stage distributions.

## Abstract

Cardiogenic shock (CS) is a life-threatening condition characterized by clinical heterogeneity and high mortality. A “one-size-fits-all” approach to management may be suboptimal. We aimed to identify distinct clinical phenotypes of CS using an unsupervised machine learning approach and to characterize their associated mortality and SCAI stages.

We conducted a retrospective analysis of 1,513 patients with CS from the Gulf registry. An unsupervised machine learning methodology was employed, using agglomerative hierarchical clustering on seven key continuous variables (Age, Ejection Fraction, Mean Arterial Pressure, Lactate, pH, Creatinine, and Alanine Transaminase) to identify patient subgroups. The optimal number of clusters was determined using a combination of quantitative metrics and clinical interpretability. The identified phenotypes were then validated against external outcomes, including in-hospital mortality and SCAI Shock Stage.

Four distinct clinical phenotypes were identified. Phenotype 1 (“Compensated Low-Risk,” n = 492, 32.5%) had the lowest mortality rate (22.4%). Phenotype 2 (“Metabolic Dysfunction,” n = 418, 27.6%) was characterized by severe left ventricular dysfunction and had a mortality of 41.9%. Phenotype 3 (“Multi-organ Failure,” n = 204, 13.5%) presented with severe metabolic, renal, and hepatic derangement and had the highest mortality (78.4%). Phenotype 4 (“Elderly Decompensated,” n = 399, 26.4%) included older patients with moderate metabolic dysfunction and had a mortality of 60.7%. A steep mortality gradient was observed across the phenotypes (p < 0.001), and the distribution of SCAI shock stages differed significantly, aligning with the risk profile of each cluster.

In a large, contemporary registry of CS patients, an unsupervised machine learning approach successfully identified four distinct and prognostically significant phenotypes. These data-driven phenotypes, characterized by unique clinical and biomarker profiles, provide a novel framework for risk stratification that moves beyond traditional classification systems and may facilitate the development of personalized therapeutic strategies for cardiogenic shock.