1347794 - Unsupervised Machine Learning Analysis of CMR-derived Dyssynchrony Metrics Identifies High-risk Fontan Phenotypes

Background: The importance of ventricular contraction patterns and dyssynchrony is poorly understood in patients with a Fontan circulation. CMR feature tracking can provide comprehensive information about the magnitude and timing of myocardial deformation. This study hypothesized that cross correlation and machine learning (ML) methods may offer superior characterization of ventricular mechanics and dyssynchrony compared to conventional analyses which emphasize peak and time-to-peak measurements.

Methods: Single-center, retrospective analysis of the first available CMR (between 2005 and 2021) for patients with a Fontan circulation, age-matched 8:1 with healthy controls. Ventricular morphology was classified as LV, RV, or mixed (when 2 ventricles contributed to systemic circulation). Feature tracking was performed on all slices of a short-axis cine stack and global and segmental data were analyzed. Twenty-four dyssynchrony metrics were calculated using circumferential and radial strain, strain rate, and displacement data. Metrics included standard deviation of time-to-peak circumferential and radial measurements for all segments (dyssynchrony index), maximum opposing wall delay, maximum base-to-apex delay, and cross-correlation analysis. K-means clustering, an unsupervised ML technique, was applied to the 24 dyssynchrony metrics standardized to z-scores for each and the optimal number of clusters was determined using an elbow technique. The primary outcome was a composite of death or listing for heart transplantation.

Results: The analytic cohort includes 512 patients (age 15y; IQR 10-21y) and 42 controls (age 16y; IQR 11-20y). Over a median follow up of 4.2y (IQR 1.7-8.8y), 58 (11.3%) patients had the composite outcome. K-means clustering reveals four clusters (Figure 1A). Cluster 4 has the highest rate of the outcome (Figure 1B) and is largely comprised of RV and mixed ventricular morphology with no controls and few patients with LV morphology. Controlling for ventricular morphology, patients in cluster 4 have an increased hazard for the primary outcome (HR = 3.6; 95% CI = 1.6-8.1; p-value = 0.002; Figure 2). Table 1 compares CMR findings between cases and controls, as well as those between cluster 4 and other clusters.

Conclusion: Unsupervised machine learning approaches, which are only trained on segmental deformation derived metrics of dyssynchrony, identify four optimal phenotypes of patients with a Fontan circulation at increasing risk for poor outcomes. Patients in the highest risk cluster are of primarily mixed or right ventricular dominant morphology and have more dilated and dyssynchronous ventricles.