Un-supervised classification of ICU patients according to congestion endotypes identifies clinical profiles associated with distinct outcomes: theCodOrea study.

Abstract Purpose In the intensive care unit (ICU), fluid overload and congestion is of daily concern. To date, congestion is defined as the linear evolution of several features, including hemodynamic and ultrasound parameters. Nevertheless, at bedside, those ultrasound features are often dissociated leading to the hypothesis that several endotypes of congestion might exist in the ICU. The aim of this study was to identify endotypes of congestion and their association to outcomes. Methods We analysed 145 patients admitted to four ICUs. Various parameters related to volume status, filling pressure, and venous congestion (fluid overload, NT-pro BNP, central venous pressure (CVP), mitral E/e' ratio, systolic/diastolic ratio of flow velocity of hepatic veins, mean inferior vena cava (IVC) diameter and its variations, stroke volume change following passive leg raising, portal vein pulsatility index, venous renal impedance) were recorded. First, unsupervised hierarchical clustering was used to identify patients endotypes. Then, we described those endotypes to allow identification of patients at bedside. Results Three distinct endotypes were identified: (1) "hemodynamic congestion" endotype (n = 75) with moderate alterations of ventricular function, increased CVP and left filling pressure, and mild fluid overload; (2) “volume overload congestion” endotype (n = 50); with normal cardiac function and filling pressure despite highly positive fluid balance (overload); (3) "systemic congestion" endotype (n = 20) with severe alterations of left and right ventricular function, increased CVP and left ventricular filling pressure values. These three sub-endotypes differed significantly by the cause of admission to ICU, the incidence of acute kidney injury, mortality and ICU/hospital length of stay. Conclusion Our un-supervised machine learning analysis identified three distinct sub- endotypes of “congestion” in ICU patients with different pathophysiologic correlates and outcomes. We also highlight key ultrasonographic features that allow identification of those endotypes at bedside.