Concealed deep phenotype abnormalities in subclinical hypertrophic cardiomyopathy detected by advanced imaging (RF_TH_276)

Predicting disease penetrance in individuals with pathogenic sarcomere mutation (G+) without hypertrophy (LVH-) is a significant clinical challenge in hypertrophic cardiomyopathy, that currently requires lifelong imaging surveillance.  We combined deep phenotyping with machine learning to detect abnormalities that may be missed by conventional tests.

Methods:

We performed a multicentre study on 155 individuals: 61 G+LVH- and 94 overt LVH+  (49 G+ / 45 G-).  All participants underwent advanced CMR with diffusion tensor imaging, quantitative perfusion and electrocardiographic imaging.

Parameters employed in unsupervised machine learning included those reflecting LVH (maximum wall thickness [MWT] analyzed using AI algorithms), ischaemia (myocardial perfusion reserve [MPR]), microstructural alteration (fractional anisotropy [FA], sheetlet orientation [E2A]), epicardial conduction (mean activation time [mean AT], gradients [G_ATmean/max] and fractionation) and repolarization (mean activation recovery interval corrected [mean ARIc], range of ARIc, gradients G_RTcmean/max).

We performed unsupervised machine learning using agglomerative hierarchical cluster analysis. Redundant CMR metrics were eliminated by removing features where pair-wise correlation was r >0.7.  The dissimilarity matrix was calculated using Euclidean distance and clusters were joined using Ward’s method, which can separate clusters even in the presence of some noise.  The optimal number of clusters was chosen using the NBclust library, which chooses the best consensus by evaluating multiple clustering validation indices.  All clustering was performed blinded to clinical data (demographics, genotype, LVH thresholds, abnormal ECG). 

Results:

Cluster validity indices calculated using advanced phenotyping implied that there were three optimal clusters (Figure).  Cluster 1 represented a mild deep phenotype (low MWT), with preserved microvascular function (higher MPR) and more preserved microstructure (higher FA, lower E2A).  Repolarization was not prolonged (lower mean ARIc) but other EP changes were similar to cluster 2 (mean AT, G_AT) and cluster 3 (Fractionation and G_RTc).

Clusters 2 and 3 displayed intermediate and severe phenotypes respectively in terms of LVH (MWT higher than cluster 1 and less than 3), ischaemia (MPR was higher than cluster 1 and less than 3), and microstructural alteration (FA was lower than cluster 1 and higher than 3).  EP changes were less severe in cluster 2 than cluster 3 in all parameters except for fractionation and G_AT.

33% of subclinical HCM clustered in intermediate or severe deep phenotypes.  These individuals had no difference in MWT and 50% had a normal ECG.

Conclusion:

A high prevalence of subclinical HCM have an intermediate or severely abnormal deep microstructural, microvascular and EP phenotype.  This subgroup had no LVH and normal 12-lead ECG, therefore are only detected by deep phenotyping techniques.