Four-dimensional left ventricular motion clustering reveals cardiovascular phenotypes at population scale

Figures

• 

  Fig. 1. Flowchart of the analysis. (a) Main steps of the generation of cardiac motion signatures and analysis of clinically relevant novel phenogroups. (b) Outline of the transformation of cardiac motion data into human-interpretable temporal signatures. Cardiovascular risk groupings and their associated cardiac motion patterns were derived from these signatures. MRI, magnetic resonance imaging; PRS, polygenic risk score; PCA, principal component analysis.

• 

  Fig. 2. Cardiac imaging segmentation and spatio-temporal models. (a) Cardiac MRI images were acquired for each participant. They capture the anatomy of the heart from base to apex of ventricles and the temporal data throughout the cardiac cycle from diastole to systole. (b) These images were then automatically segmented with our algorithm to locate the right ventricle and the left ventricle myocardium and blood pool. (c) These segmentation masks were then fitted and registered to a common atlas space between participants to develop a consistent three-dimensional model of the left ventricle. (d) Deformation fields were used to generate the entire cardiac cycle of surface meshes.

• 

  Fig. 3. Convolutional variational autoencoder (CVAE) and temporal signatures. The cardiac cycle was fed in the model as a collection of time windows. The encoder and decoder were made of two-dimensional convolution layers. The encoder took as input the collection of windows and encoded a latent distribution of mean μ and standard deviation σ. The latent space z was sampled from that distribution and was fed back into the model through the decoder. The decoder aimed to reconstruct the input cardiac cycle. The reconstruction loss between input and model reconstruction was assessed using a mean squared error (MSE). The model was trained with the weighted sum of MSE and Kullback–Leibler (KL) divergence, which ensured the coherent continuity of the latent space. After training, the dimensionality of the latent space was further reduced with principal component analysis (PCA). Subsequent windows were concatenated and visualised in the reduced space in the form of temporal signatures describing the input cardiac cycle. The detailed layer-by-layer architecture of the CVAE is provided in Supplementary Fig. 5.

• 

  Fig. 4. Signature distribution by feature z-score and signature average by binary outcome, on latent space position. (a) Temporal signatures are plotted for relevant variables grouped by increasing z-scores. For each variable, individuals were stratified into five evenly spaced z-score bins within the range [-2,2], and the mean latent signature was computed for each group. Feature variations shift signatures from the population average signature, in black on each plot. (b) The average signature is plotted for each cardiovascular risk, prevalent and incident outcome. HDL, high-density lipoprotein; HbA1c, glycated hemoglobin; PRS, polygenic risk score; DCM, dilated cardiomyopathy; HCM, hypertrophic cardiomyopathy; LV, left ventricle; EF, ejection fraction; WT, wall thickness; MRI, magnetic resonance imaging.

• 

  Fig. 5. Combination of clustering on in-plane position and contour. Temporal signatures were clustered in two stages. A few individual examples are shown here to simplify the figure. First, the signatures were centred around the origin and clustered on their relative coordinates, capturing the contour of the trajectories, resulting in two clusters. Second, clustering was done on the original, uncentred coordinates of the temporal signatures, capturing their in-plane position, and revealing three clusters. The results of both methods were combined to give a dual clustering of the latent space, giving six clusters numbered from 1 to 6.

• 

  Fig. 6. Clinical interpretation of motion-based phenogroups. (a) Barycentre position of all individual trajectories, coloured by their assigned phenogroups, in the latent space of the two principal components. (b) Average signatures for each phenogroup centred around the origin of the latent space. (c) Summary of findings characterising the cardiovascular risks and disease status of the phenogroups. Participant characteristics and diagnostics were used to highlight the differences between groups. Findings are reported as the median confidence interval for continuous variables (Fig. 7) and odds ratio and direction of association for discrete features (Fig. 8) following statistical testing. A Kruskal-Wallis test followed by a post-hoc Dunn test was used for continuous variables, while a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\chi ^{2}$$\end{document} test followed by a post-hoc Fisher test was performed for discrete variables. Multiple testing was accounted for by the Benjamini-Hochberg procedure, and…

• 

  Fig. 7. Numerical features significantly associated with motion-based phenogroups. Numerical features were tested for significant associations with the phenogroups. A Kruskal-Wallis test was performed, followed by a post-hoc Dunn test revealing pairs of phenogroups sharing significant differences in variable distribution. Benjamini-Hochberg correction for multiple testing was done, taking into consideration discrete features as well (see Supplementary Data for all tested features). The significance level was P < 0.05 and the sample size was n = 20,318. The median of the z-score and its confidence interval are displayed for each phenogroup. For a given feature, groups share a letter if there is no significant difference between them for that feature. MRI, magnetic resonance imaging; DCM, dilated cardiomyopathy; PRS, polygenic risk score; HT, hypertension; LDL, low-density lipoprotein; T/S, telomere to single copy gene; HCM, hypertrophic cardiomyopathy; LV, left ventricle; ED, end-diastolic.

• 

  Fig. 8. Discrete and binary features significantly associated with motion-based phenogroups. Discrete features were tested for significant associations with the phenogroups. A \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\chi ^{2}$$\end{document} test was used, followed by a post-hoc Fisher’s unit test which highlighted significant enrichment for discrete features. Benjamini-Hochberg correction for multiple testing was done, taking into consideration numerical features as well (refer to Data for all tested features and detailed statistical results). The significance level was P < 0.05 and the sample size was n = 20,318. The size of the triangle reflects the logarithm of the P value obtained with each statistical test while its direction is of the association. The odds ratio is also displayed for each significant association.

• 

  Fig. 9. Left ventricle motion velocity by phenogroup - inferolateral view. (a) Zoomed in examples of the left ventricle during systole and diastole. The velocity of the ventricle wall is displayed on the surface, relative to its barycentre. The velocity is positive during relaxation, or diastole when the ventricle is filling with blood, while it is negative during contraction, or systole when ventricular ejection occurs. Red beams indicate areas of faster movement relative to the population average motion, while blue beams show areas of relative slower movement. (b) The left ventricle motion was compared to the population’s average motion for each phenogroup. The six frames represent one entire cardiac cycle from systole (contraction of the ventricle) to diastole (relaxation of the ventricle). The velocity of the ventricle was calculated as the local gradient of change between two consecutive frames, averaged out over multiple frames to reduce the number of displayed frames, and normalized for the size of the ventricle.