Stability-driven phenotype clustering of children with hypertrophic cardiomyopathy identifies phenogroups with distinct outcomes

Introduction The aetiology, phenotype and natural history of childhood-onset hypertrophic cardiomyopathy (HCM) are highly variable, and it remains challenging to predict outcomes using traditional statistical methods. This study used stability-driven phenotypic clustering methods to identify robust and reproducible subgroups of patients with distinct disease trajectories. Methods Retrospective clinical phenotyping data of 15 variables were collected from baseline clinical assessment for patients (≤18 years) with HCM at a single quaternary referral centre. A stability-driven optimisation process over unsupervised clustering methods was used to select the most robust method on the training dataset and demonstrated generalisability on an internal test set, with concordance measured by the Adjusted Rand Index (ARI). Final clusters were created with bootstrapping and consensus clustering on the full development cohort. The association between phenotypic clusters and clinical outcomes (death /heart transplant, arrhythmia, or heart failure event) were analysed. Findings were validated in an external independent validation cohort from the International Paediatric Hypertrophic Cardiomyopathy Consortium (IPHCC) (n=1954). Results In the development cohort of 512 patients [median age 7.0(1.0 - 13.0), n= 63% male], the optimal clustering method was k-means with 3 clusters (table 1). Cluster A was characterised by the youngest patients (median age 1.0 (1.0 - 4.0), p ≤0.001), with one third having variants in RASopathy genes. Cluster B had severe hypertrophy [z- maximal wall thickness (MWT) maximal wall thickness (MWT) 17.6 (14.7 - 22.0)] with the highest prevalence of obstruction (n=74, 64.9%, p≤0.001) and symptoms. Cluster C were older [13.0 (11.0 - 15.0), 60% female, and had milder hypertrophy [zMWT 6.5 (4.4 - 10.0)] with a low prevalence of obstruction (n= 3.7%). These three phenotypic clusters were associated with distinct event-free survival. The incidence of arrhythmic events was highest in cluster B (n= 20% within 5 years), whilst cluster C had the lowest incidence of mortality/heart transplant (n= 4% within 5 years) (Figure 1a). A decision tree trained on the development cohort using eight clinical variables (age, height, weight, MWT, Left Ventricular Internal Dimension in Diastole (LVIDD), Left Ventricular Outflow Tract (LVOT) rest, Interventricular Septal Thickening (IVST) and chest pain) was able to classify patients from the external validation cohort into the three phenotypic clusters that remained associated with event-free survival (Figure 1b). Conclusions This study identifies three clusters of childhood HCM patients associated with clinically important differences in baseline characteristics and long-term outcomes. This demonstrates that stability-driven machine learning techniques could be a valuable tool for risk stratification even in rare diseases with small datasets.Figure 1Table 1