Phenotypic Clustering of Beta-Thalassemia Intermedia Patients Using Cardiovascular Magnetic Resonance

Abstract

We employed an unsupervised clustering method that integrated demographic, clinical, and cardiac magnetic resonance (CMR) data to identify distinct phenogroups (PGs) of patients with beta-thalassemia intermedia (β-TI). We considered 138 β-TI patients consecutively enrolled in the Myocardial Iron Overload in Thalassemia (MIOT) Network who underwent MR for the quantification of hepatic and cardiac iron overload (T2* technique), the assessment of biventricular size and function and atrial dimensions (cine images), and the detection of replacement myocardial fibrosis (late gadolinium enhancement technique). Three mutually exclusive phenogroups were identified based on unsupervised hierarchical clustering of principal components: PG1, women; PG2, patients with replacement myocardial fibrosis, increased biventricular volumes and masses, and lower left ventricular ejection fraction; and PG3, men without replacement myocardial fibrosis, but with increased biventricular volumes and masses and lower left ventricular ejection fraction. The hematochemical parameters and the hepatic and cardiac iron levels did not contribute to the PG definition. PG2 exhibited a significantly higher risk of future cardiovascular events (heart failure, arrhythmias, and pulmonary hypertension) than PG1 (hazard ratio-HR = 10.5; p = 0.027) and PG3 (HR = 9.0; p = 0.038). Clustering emerged as a useful tool for risk stratification in TI, enabling the identification of three phenogroups with distinct clinical and prognostic characteristics.

Keywords: cardiovascular magnetic resonance imaging; clustering; phenomapping; thalassemia intermedia.

Figure 1

Cluster analysis. (A) Biplot with the first two components from the principal component analysis. Points represent individual patients displayed based on their individual characteristics. The colors represent the clear differentiation of the three phenogroups from the cluster analysis. (B) Dendrogram of the hierarchical cluster analysis. The ordinate axis represents the distance between merged phenogroups. Dashed lines indicate the partitioning with three phenogroups.

Figure 2

Characteristic plots of the three mutually exclusive phenogroups, including their most representative clinical variables. The over- or underrepresentation of a variable within a cluster was analyzed by a v-test within the hierarchical clustering of principal component function based on the hypergeometric distribution. A positive value indicates that the variable is overrepresented in the relevant phenogroup, while a negative value implies an underrepresentation of the corresponding variable. Only significant variables with p < 0.05 are listed.

Figure 3

Variable importance plot from the supervised random forest model. Variables appear in decreasing order based on the contribution to the differentiation of patients in phenogroups.

Figure 4

(A) Frequency of cardiovascular events in the three phenogroups. (B) Kaplan–Meier curves for cardiovascular events stratified by phenogroup.