Human phenotype ontology annotation and cluster analysis for pulmonary atresia to unravel clinical outcomes

Abstract

Background: Pulmonary atresia (PA) is a heterogeneous congenital heart defect and ventricular septal defect (VSD) is the most vital factor for the conventional classification of PA patients. The simple dichotomy could not fully describe the cardiac morphologies and pathophysiology in such a complex disease. We utilized the Human Phenotype Ontology (HPO) database to explore the phenotypic patterns of PA and the phenotypic influence on prognosis.

Methods: We recruited 786 patients with diagnoses of PA between 2008 and 2016 at Fuwai Hospital. According to cardiovascular phenotypes of patients, we retrieved 52 HPO terms for further analyses. The patients were classified into three clusters based on unsupervised hierarchical clustering. We used Kaplan-Meier curves to estimate survival, the log-rank test to compare survival between clusters, and univariate and multivariate Cox proportional hazards regression modeling to investigate potential risk factors.

Results: According to HPO term distribution, we observed significant differences of morphological abnormalities in 3 clusters. We defined cluster 1 as being associated with Tetralogy of Fallot (TOF), VSD, right ventricular hypertrophy (RVH), and aortopulmonary collateral arteries (ACA). ACA was not included in the cluster classification because it was not an HPO term. Cluster 2 was associated with hypoplastic right heart (HRH), atrial septal defect (ASD) and tricuspid disease as the main morphological abnormalities. Cluster 3 presented higher frequency of single ventricle (SV), dextrocardia, and common atrium (CA). The mortality rate in cluster 1 was significantly lower than the rates in cluster 2 and 3 (p = 0.04). Multivariable analysis revealed that abnormal atrioventricular connection (AAC, p = 0.011) and persistent left superior vena cava (LSVC, p = 0.003) were associated with an increased risk of mortality.

Conclusions: Our study reported a large cohort with clinical phenotypic, surgical strategy and long time follow-up. In addition, we provided a precise classification and successfully risk stratification for patients with PA.

Keywords: Cox proportional hazards regression; Human Phenotype Ontology; Kaplan-Meier curves; pulmonary atresia; unsupervised cluster analysis.

FIGURE 1

Clinical characteristics of the confirmed PA patients. (A) The number of PA patients in different age and gender groups, neonate (age < 1), infant (1 ≤ age ≤ 2), child (3 ≤ age ≤ 12), and adolescent (13 ≤ age < 18). (B) The number of PA patients undergone different types of surgery.

FIGURE 2

HPO terms annotated for the clinical phenotype of PA patients. (A) The ontology plot presented the relationship of all HPO terms. The circles with border showed the all phenotypes in PA patients. The color indicated the frequency of terms in the HPO database. The arrows were “is-a” relations between terms in the ontology. (B) The distribution of the number of HPO terms per index case. (C) The frequency distribution of HPO terms in our cohort.

FIGURE 3

Heatmap of the phenotypic similarity clustering of PA patients. The heatmap was calculated by unsupervised clustering from the distance matrix of the phenotypic similarity of all patients. The line of dashes represented the height to cut the tree into three clusters. The color indicated the similarity between patients.

FIGURE 4

(A) Kaplan-Meier curve for clinical outcomes of three clusters. The Kaplan-Meier curve indicated the survival rate of cluster 1 was significant higher than cluster 2 and 3 (p = 0.04). (B) The age of death of the patients in cluster 3 was significantly higher than that of the cluster 1 and 2.