ID:65 FC-D11-8 Unsupervised machine learning identifies distinct phenotypes of cardiogenic shock undergoing VA-ECMO with differential prognosis

Background: Cardiogenic shock (CS) is a life-threatening condition marked by inadequate cardiac output and systemic hypoperfusion. Despite widespread use of venoarterial extracorporeal membrane oxygenation (VA-ECMO) for CS, outcomes remain heterogeneous. We sought to identify clinically meaningful phenotypes and assess their prognostic value in patients receiving VA-ECMO. Methods: In this retrospective study of 135 patients with CS treated with VA-ECMO, unsupervised machine learning using factor analysis of mixed data followed by clustering was applied to baseline clinical, hemodynamic, and biochemical variables. Incremental prognostic value was evaluated by comparing Cox models with and without phenotype assignment, adjusting for the SAVE score, vasoactive-inotropic score (VIS), and prior medical history. Discrimination was assessed using Harrell’s C-index with 1,000-bootstrap validation. Results: Three phenotypes were identified. Phenotype 3 (n=29) exhibited severe hemodynamic instability, higher VIS, and marked hyperlactatemia. After optimism correction, the phenotype-enhanced model achieved a C-index of 0.706 (95% CI 0.684–0.719) versus 0.677 (95% CI 0.654–0.686) for the baseline model (ΔC-index 0.029). Phenotype 3 was independently associated with increased 28-day mortality (adjusted hazard ratio 3.36, 95% CI 1.50–7.54), whereas phenotype 2 did not differ from phenotype 1. Conclusion: Machine learning–derived phenotypes stratify mortality risk in CS patients receiving VA-ECMO and provide incremental prognostic information beyond established scores.