Machine learning to phenotype hemodynamic response to va ecmo and clinical outcomes in cardiogenic shock

Background

Simulation studies suggest that veno-arterial extracorporeal membrane oxygenation (VA ECMO) support in cardiogenic shock (CS) increases afterload, resulting in left ventricular (LV) distension. We propose that the effect of VA ECMO on the LV is heterogenous, because VA ECMO have discordant effects on LV input (transpulmonary blood flow) and LV output (stroke volume).

Purpose

Machine learning to phenotype hemodynamic response to VA ECMO, and examine outcomes related to hemodynamic phenotype.

Method

This study has 2 parts. Firstly, changes in LV end-diastolic pressure (LVEDP) and arterial pulse pressure (PP) were assessed in 58 pigs with CS (coronary occlusion/ global myocardial hypoxia) from baseline (<1L/min) to 2L/min of VA ECMO flow. Secondly, 128 patients with CS due to acute myocardial infarction, myocarditis or end-stage heart failure were phenotyped using hemodynamic data within 2 hours of VA ECMO. K-means clustering included 3 features - pulmonary artery diastolic pressure (PADP surrogate for LV distension), PA pulse pressure (PAPP, surrogate for transpulmonary blood flow) and arterial PP (surrogate for LV output). Outcomes were 12-hour lactate clearance and in-hospital mortality.

Results

In the animal study, increasing VA ECMO flow was associated with increased LVEDP from 13.8±4.7, to 14.6±5.7 mmHg (p=0.011), and decreased arterial PP (29.0±13.2 to 25.3±12.3 mmHg, p<0.001), but individual responses were variable. K-means clustering identified 3 clusters from delta LVEDP and delta arterial PP [Fig 1 top].

Three phenotypes were identified from the 128 patients with CS [Fig 1 bottom] - the 'pulsatile' phenotype (n=34, 26%) had the highest arterial PP (30.9±5.1 vs 18.3±3.7 vs 13.4±4.9, p<0.001), the 'low pulsatility' phenotype (n=66, 52%) had low arterial PP and PA PP (8.2±1.8 vs 21.8±2.8 vs 18.7±2.4, p<0.001), and the 'LV distension' phenotype (n=28 or 22%) had elevated PADP (29.8±2.8 vs 18.2±3.1 vs 17.2±2.4, p<0.001). Acute MI was more common in 'LV distension' phenotype and HF-related CS more common in 'low pulsatility' phenotype (Bonferroni p<0.05). The 'low pulsatility' phenotype had more severe RV dysfunction: lower TAPSE (10 (9-12) vs 13 (11-14) vs 13(12-15)mm), larger RV (mid-RV diameter 3.9 (3.6-4.4) vs 3.2 (2.9-3.5) vs 3.1 (2.9-2.5)cm) and more severe TR (27% vs 3% and 7%) (all p<0.05). The 'Pulsatile' phenotype had the highest 12-hour lactate clearance (58% (41-70) vs 48 (37-54) vs 43 (32-56), Bonferroni p<0.05). The 'pulsatile' phenotype had the lowest in-hospital mortality (29% vs 47% vs 50%, p<0.05) [Fig 2].

Conclusion

Animal and human data showed heterogenous hemodynamic response to VA ECMO related to differences in LV inflow and outflow (i.e. the 'pulsatile', 'low pulsatility' and 'LV distension' post-VA ECMO phenotypes). The phenotypes were related to the underlying CS etiology and right heart function . The 'pulsatile' phenotype had the highest 12-hour lactate clearance and in-hospital survival.Fig 1:Clustering analysisFor image description, please refer to the figure legend and surrounding text.Fig 2:Outcomes by etiology & phenotypeFor image description, please refer to the figure legend and surrounding text.