Unsupervised Machine Learning of LGE Patterns on Cardiac MRI Identifies Patients at Risk for Right Ventricular Failure After LVAD

Introduction Right ventricular failure (RVF) is a major cause of morbidity and mortality after left ventricular assist device (LVAD), however predicting RVF remains challenging. Hypothesis We hypothesized that an analysis of late gadolinium enhancement (LGE) patterns on pre-operative cardiac MRI (cMRI) could identify patients at risk for RVF after LVAD. Methods We analyzed reports for cMRIs performed on patients within one year prior to LVAD at our institution and abstracted LGE patterns using the 17-segment model. Patients were then grouped into clusters by similarities in LGE patterns using an unsupervised machine learning (ML) algorithm of hierarchical agglomerative clustering. Statistical comparison of the resulting clusters was then performed. Results Patients (N=31) were grouped into 3 clusters (Figure) with varying patterns of LGE. Cluster 1 patients (n=16) had no LGE or atypical LGE patterns and were significantly younger (age 42 ± 18) than other clusters (p=0.029). Cluster 2 patients (n=11) had extensive transmural LGE patterns and were more likely to have hypertension (p=0.006) and dyslipidemia (p=0.002) than other groups. Cluster 3 patients (n=4) had some degree of subendocardial LGE but no extensive transmural LGE patterns. No patients in cluster 2 developed RVF after LVAD, while 4 patients (25%) in cluster 1 and 2 patients in cluster 3 (50%) had RVF after LVAD, though the difference between groups did not reach statistical significance due to small number of patients in the cohort overall (p=0.058). Importantly, traditional factors associated with RVF including hemodynamics and echocardiographic/MRI parameters of LV and RV function were not significantly different between clusters. Further, LGE enhancement of the RV myocardium or RV insertion points were not associated with RVF after LVAD. Conclusions Unsupervised ML of LGE patterns on cMRI can identify clusters of patients at risk for RVF. LGE patterns on cMRI may identify patients with non-ischemic (cluster 1) and mixed (cluster 3) etiologies of their heart failure who are at higher risk for developing RVF due to global biventricular myocardial involvement than patients with a truly ischemic etiology of their heart failure (cluster 2). Future research in a larger cohort is needed to confirm this hypothesis. Figure: Dendrogram produced from agglomerative hierarchal clustering of LGE analysis of cardiac MRIs using standardized myocardial segmentation