Unsupervised CMR Phenotyping of Hypertrophic Cardiomyopathy

Continuous variables are presented as mean ± standard deviation and compared using either parametric t-tests or non-parametric Mann–Whitney U tests, as appropriate. Categorical variables are reported as frequencies with corresponding percentages and compared using the chi-squared test or Fisher’s exact test, where applicable.Abbreviations: BMI – body mass index; ICD – implantable cardioverter-defibrillator; NYHA – New York Heart Association; LVOT – left ventricular outflow tract; ACE – angiotensin-converting enzyme; ARB – angiotensin receptor blocker; LV/RV – left/right ventricle; LA – left atrium; WT – wall thickness; MI – mass index; EDVI/ESVI – end-diastolic/end-systolic volume index; SVI – stroke volume index; EF – ejection fraction; LGE – late gadolinium enhancement.Figure 1: Preprocessing of extracted features and unsupervised consensus clustering. A total of 23 features per patient (9 demographic and clinical variables, 14 CMR-derived features) were extracted. Missing data were imputed separately for numerical variables using the MissForest algorithm (based on a random forest model), and for categorical variables using multiple imputation by chained equations (MICE) with logistic regression models. The maximum missingness rate among features was 5.5%. Gower’s unified distance was then computed between individual patients by calculating differences in numerical and categorical features separately and combining them into a single distance metric. The resulting Gower distance matrix was resampled 100 times, each containing 80% of patients. Each of the 100 subsets was clustered using partitioning around medoids (PAM), and pairwise results were summarized in a consensus matrix to determine the optimal number of clusters. Abbreviations: BSA – body surface area; CAD – coronary artery disease; NYHA – New York Heart Association; AF – atrial fibrillation; LVOT – left ventricular outflow tract; LA – left atrium; LGE – late gadolinium enhancement.
Figure 2. Patient clusters and adverse event rates across clusters in the development and testing set. Figure 2A and 2C display heatmaps of feature distributions across the identified clusters in the development and testing set. For numerical features, standardized Z-scores relative to the cohort mean are shown, with lower values (0 to -1) represented in shades of blue and higher values (0 to +1) in shades of red. For categorical features, the within-cluster frequency of each feature is visualized, with warmer colors indicating a higher proportion (%) of patients exhibiting that characteristic. Figure 2B and 2D show Kaplan–Meier survival curves for heart failure events, new-onset atrial fibrillation, and sudden cardiac death across the three clusters in the development and testing set, demonstrating differences in event rates for all outcomes between clusters. Abbreviations: BSA – body surface area; CAD – coronary artery disease; LV/RV – left/right ventricle; LA – left atrium; MI – mass index; EDVI/ESVI – end-diastolic/end-systolic volume index; SVI – stroke volume index; EF – ejection fraction; LGE – late gadolinium enhancement; WT – wall thickness; LVOT – left ventricular outflow tract; NYHA – New York Heart Association