Multidimensional clinical phenotyping of an adult cystic fibrosis patient population

Abstract

Background: Cystic Fibrosis (CF) is a multi-systemic disease resulting from mutations in the Cystic Fibrosis Transmembrane Regulator (CFTR) gene and has major manifestations in the sino-pulmonary, and gastro-intestinal tracts. Clinical phenotypes were generated using 26 common clinical variables to generate classes that overlapped quantiles of lung function and were based on multiple aspects of CF systemic disease.

Methods: The variables included age, gender, CFTR mutations, FEV1% predicted, FVC% predicted, height, weight, Brasfield chest xray score, pancreatic sufficiency status and clinical microbiology results. Complete datasets were compiled on 211 subjects. Phenotypes were identified using a proximity matrix generated by the unsupervised Random Forests algorithm and subsequent clustering by the Partitioning around Medoids (PAM) algorithm. The final phenotypic classes were then characterized and compared to a similar dataset obtained three years earlier.

Findings: Clinical phenotypes were identified using a clustering strategy that generated four and five phenotypes. Each strategy identified 1) a low lung health scores phenotype, 2) a younger, well-nourished, male-dominated class, 3) various high lung health score phenotypes that varied in terms of age, gender and nutritional status. This multidimensional clinical phenotyping strategy identified classes with expected microbiology results and low risk clinical phenotypes with pancreatic sufficiency.

Interpretation: This study demonstrated regional adult CF clinical phenotypes using non-parametric, continuous, ordinal and categorical data with a minimal amount of subjective data to identify clinically relevant phenotypes. These studies identified the relative stability of the phenotypes, demonstrated specific phenotypes consistent with published findings and identified others needing further study.

Fig 1. Comparison of the 2011 and 2014 adult CF patient cohorts.

(A) Notchplots comparing the FEV1% predicted, FVC% predicted, Age*FEV1% product, Brasfield chest Xray scores, age and BMI. (B) Proportions of pancreatic sufficiency (PS), fraction of male subjects (f Male) and presence of two CFTR Class I, II and III mutations. The Fisher exact test odds ratio (OR) with the p-value are presented. Notchplots demonstrate the median, 25 percentile, 75 percentile and outlier values. The extending bars demonstrate the span between 1.5 x interquartile range above and below the median. Nonoverlapping notched areas likely represent significant differences between two groups.

Fig 2. Characterization of the Adult CF Clinical Classes.

(A) The variable importance plot generated by the supervised Random Forests algorithm of the classes derived from the proximity matrix of the unsupervised Random Forests (all variables) and the PAM clustering algorithm (k = 5 is shown). (B) The confusion matrix generated from the supervised Random Forests algorithm of the classes that were reformed after the described dimension reduction strategy. Overall out of bag error rates were 8.06% and 11.4% for the k = 4 and k = 5 classes, respectively.

Fig 3. Visualization of the class separation (k = 4 and k = 5) using multidimensional scaling.

(MDS) of the proximity matrix generated by the eight important clinical variables.

Fig 4. Multidimensional clinical phenotype (MDCP) descriptions.

(A) Mean values of the FEV1%, total Brasfield chest xray score, age, age*FEV1% predicted, body mass index (BMI) and the fraction of males in each phenotype are shown. (B) Notchplots of the FEV1% predicted, FEV1 FVC ratio, age*FEV1% predicted, Brasfield chest xray score, age and BMI are plotted. The color of the notchplot boxes indicate the proportion of males in each class ranging from 0.0 male (green),. 5 male (white) to 1.0 male (blue).

Fig 5. Clinical Phenotype Transitions.

(A) The locations of subjects with deaths/transplants and class transitions are plotted on the MDS plot generated from the 2014 data (light gray dots). The colored points show the location and classes of patients present in both datasets. Positions with central black dots indicate subjects who died or had a lung transplant. Positions with thick black outlines of the points demonstrated class transitions. (B) Counts and frequencies of transitions between two phenotypes of subjects present in both cohorts (k = 5 phenotypes). The table shows the proportion of subjects leaving a particular clinical phenotype (p Transitions) and the fraction of total class transitions into a given clinical phenotype (f Transitions). (C) The counts of deaths or lung transplants in subjects represented in the 2014 cohort (k = 4 and k = 5) class phenotypes.