Skip to main content
2025 SCMR Conference Main banner
Final Program


Schedule at a Glance
Back
On-Demand Video

Science Sessions

Science Session 3: Artificial Intelligence in CMR

Phenotypic clustering of hypertrophic cardiomyopathy using 3D cardiac shape models and associations with future adverse outcomes

Friday, January 31, 2025
8:40 AM – 8:50 AM East Coast USA Time
Location: Blue Room
Background: Current risk scores for the prediction of future adverse events in patients with hypertrophic cardiomyopathy (HCM) incorporate the severity of myocardial hypertrophy, described as a single measure of the maximum left ventricular (LV) wall thickness. This manual 2D measurement is recognized to be an inaccurate, crude marker of disease expression and sub-optimal risk predictor. In this study, we aimed to (i) identify novel phenotypic clusters of HCM based on CMR-based 3D cardiac shape models presenting detailed assessments of regional wall thickness and (ii) study the associations of these phenotypic clusters with future adverse outcomes.

Methods: 405 patients with CMR-confirmed diagnosis of HCM were identified from the Cardiovascular Imaging Registry of Calgary (CIROC). All patients completed baseline questionnaires and a standard CMR imaging protocol inclusive of cines and were followed for a composite outcome of all-cause mortality, heart failure hospitalization, ventricular arrhythmia, appropriate implantable cardioverter-defibrillator therapy, new-onset atrial fibrillation, and survival of sudden cardiac arrest. CMR image processing was performed by segmentation of long and short-axis cines, followed by image alignment, registration, and bi-layer meshing of the LV. The latter generated a 3D shape model with 960 spatially registered nodes with triangular faces for both endocardial and epicardial layers, arranged into 456 hexahedral elements (Figure 1A and 1B). Element thicknesses were extracted, followed by dimensionality reduction using principal component analysis (PCA). The first 8 PCA components were fed into a k-means unsupervised machine learning clustering algorithm, yielding 2 clusters (Figure 1C) that were compared with respect to baseline clinical and imaging characteristics, as well as future adverse outcomes.

Results: Characteristics of the two clusters are shown in Table 1. Compared to Cluster 1 (252 patients; 62%), Cluster 2 were more likely males, with a higher prevalence of diabetes and hypertension, as well as higher chamber volumes, LV mass, and LV scar/fibrosis burden. Over a median follow-up of 4.5 years, 67 patients developed the composite outcome (11% in Cluster 1 vs 25% in Cluster 2; p< 0.001). Kaplan-Meier curves for association of cluster membership with the outcome are shown in Figure 1D. Adjusting for age, sex, left atrial volume, LV mass, and LV fibrosis/scar burden, Cluster 2 was independently associated with worse outcomes (adjusted HR 2.2 [95% CI 1.2 - 4.3]; p = 0.013, Figure 2).

Conclusion: Using unsupervised machine learning applied to comprehensive regional LV wall thickness measurements generated from a 3D mesh model, we identified two unique phenotypic clusters of HCM. Cluster membership was independently associated with future adverse clinical outcomes. Future studies are warranted to confirm the value of this approach in refining currently existing scores for risk stratification of patients with HCM.

(A) Example of segmentations of the left ventricle (LV) from long and short axis cines used to generate (B) a 3D bi-layer mesh model of the LV. (C) 3D plot of the two clusters resulting from k-means clustering algorithm applied to the first eight components of principal component analysis (PCA). Each colour (red and blue) represents a unique cluster. Plot axes represent the first three PCA components. (D) Kaplan-Meier survival curves for association of cluster membership with future adverse outcomes.

Forest plot of the results of the multivariable Cox model for associations of cluster membership with future adverse outcomes.