As HIV-infected infants live longer, it is increasingly evident that HIV or certain antiretroviral therapies can cause distorted body shapes with excess fat in the belly or loss of fat in the extremities, often accompanied by abnormal lipids and glucose metabolism. To better understand the prevalence of these abnormalities during childhood and adolescence, the Pediatric AIDS Clinical Trials Group conducted a cross-sectional study, which enrolled HIV-infected and HIV-uninfected participants (P1045).1 An array of outcomes was measured, including tests of metabolic function, body composition, lifestyle, and diet. To date, analyses have focused on single outcomes,1,2 but an additional goal of the study was to understand the relationship among groups of outcomes and whether these relationships varied by factors such as HIV status, pubertal development, sex, and race/ethnicity.
Cluster heat maps3 are an easy and effective way of visually displaying multivariate (multiple outcome) data, combining the use of color to distinguish the magnitude of measurements and dendrograms (tree-like figures showing a hierarchical grouping) to show clustering4 of individuals and outcome measurements. Heat maps are often used for DNA microarray data to identify which genes (among hundreds or thousands) are associated with disease. In this article, we use heat maps to investigate clustering of measurements of body shape, lipids, and glucose metabolism and to see whether these patterns are related to HIV status and other variables not used to generate the heat map.
P1045 enrolled 240 HIV-infected participants group-matched by Tanner stage, sex, and race/ethnicity to 146 HIV-uninfected participants.1 The institutional review board at each clinical site approved the study, and appropriate informed consent was obtained before enrollment. Participants ranged in age from 7 to 24 years (median age 12.4 years, 55% male, 55% African American, 11% white). Dual energy x-ray absorptiometry was performed to estimate total body mass, total body fat (TBF), extremity (arm + leg) fat (EXF), trunk fat, leg mass, and leg fat (LEGF), with results available on 379 of the 386 participants (236 HIV infected and 143 HIV uninfected). Fasting lipids [triglycerides, total and high-density lipoprotein (HDL) cholesterol], glucose and insulin were measured and high-sensitivity C-reactive protein (CRP), which is a measure of inflammation and may be predictive of future cardiovascular disease (n = 343). The following outcome measures were calculated for this analysis: percent body fat (% body fat) = TBF/total body mass; percent LEGF (% LEGF) = LEGF/leg mass; trunk-to-limb fat ratio = trunk fat/EXF; total cholesterol to HDL ratio, and the homeostasis model of insulin resistance (HOMA-IR5). Each of these outcomes depends to a varying extent on age, sex, and race/ethnicity, and they differ in magnitude and variability. To put all outcomes on a similar scale, each was standardized into age and sex-adjusted z scores using the HIV uninfected as the reference population.1
Using the standardized outcomes, a heat map was generated using the heatmap.2 function in the gplots library of R.6 The heat map shows dendrograms (hierarchical tree) for participants on one axis and for the outcomes on the second axis. The vertical length of the branches in the dendrogram represents the degree of separation between individuals or outcomes. Clustering was done using Ward's method, which clusters by minimizing the sum of squared deviations of each point from the mean of its cluster, and which tends to result in spherical clusters.4 Three colors were used in the body of the heat map representing the magnitude of outcome values with z scores <−1 in yellow, −1 ≤ z ≤ 1 in blue, and z >1 in red. The distribution of HIV-infected (green) and HIV-uninfected (blue) participants across clusters is shown along the top of the heat map. Choice of the best number of clusters was subjective and based loosely on the vertical length of the branches in the dendrograms. After selecting the number of clusters at the participant level, we calculated the median z score for each outcome variable within each cluster.
To evaluate associations between clusters of participants with HIV status, sex, Tanner stage, race/ethnicity, and CRP (variables not included in generating the heat map), χ2 tests were used in univariate analyses and multinomial regression for multivariate analyses with cluster as the response variable and including all covariates.7 To assess robustness of conclusions, 500 datasets were generated where participants were randomly assigned to a cluster. A χ2 P value was calculated for each cross tabulation of cluster membership with each external variable. This created a null distribution of P values against which the observed P value was compared (the permutation P value). Bootstrapping was used to check for consistency of cluster phenotype attributes across samples.
To evaluate associations between high CRP levels and phenotype (cluster), sex, Tanner stage, and race/ethnicity, logistic regression models were fit with abnormal CRP as the outcome and with cluster and each of the other 3 covariates one at a time in HIV-infected and HIV-uninfected participants separately.
P values less than 5% were used for highlighting statistical significance, but these analyses were purely exploratory.
The heat map in Figure 1 illustrates the dendrogram for the 7 outcome variables on the horizontal axis and the dendrogram for participants on the vertical axis. There were 2 distinct clusters of outcome variables at the highest level of separation. The first cluster (A) contains percent LEGF, percent TBF, and HDL cholesterol; and the second cluster (B) contains HOMA-IR, trunk to limb fat ratio, cholesterol to HDL ratio, and triglycerides.
For participants, there were 3 major clusters named for body phenotype (Table 1). Compared with the other 2 clusters, cluster 1 (large) had the highest median z scores for body composition outcomes (% LEGF z = 0.64, % body fat z = 0.84, trunk to limb ratio z = 0.60), and glucose metabolism (HOMA-IR z = 0.58) and second highest medians for lipids (cholesterol to HDL ratio z = 0.71, triglycerides z = 0.95 and HDL z = −0.70). This is seen in the heat map as mostly yellow coloration for HDL (for which higher values are better), and blue and red coloration across the other 6 outcomes. Phenotypically, these were subjects with above average percent fat exhibiting evidence of dyslipidemia and greater insulin resistance. Cluster 2 (thin) had the highest median triglycerides (z = 2.27) and cholesterol to HDL ratio (z = 1.38), the lowest median HDL (−0.71), % total fat (−0.89) and % LEGF (−1.22) and moderate to high trunk to limb ratio (0.23). The low percent fat and moderate to high trunk to limb ratio suggest loss of EXF along with dyslipidemia, but normal glucose metabolism. Cluster 3 (average) had the highest median HDL (z = 0.22) and the lowest median triglycerides (z = −0.31), cholesterol to HDL (z = −0.41), trunk to limb ratio (z = −0.30), and HOMA-IR (z = −0.45), representing more normal body phenotype, lipids, and glucose metabolism.
We then investigated whether membership in 1 of the 3 participant clusters was related to HIV status, sex, Tanner stage, race/ethnicity or CRP. Proportions of participants by each characteristic within each cluster are shown in Table 1. In univariate analyses evaluating each predictor separately, only HIV status (P < 0.001) and CRP (P < 0.001) were related to cluster membership. The high proportion of HIV-infected participants in cluster 2 (thin, 90.9%) as compared with cluster 1 (large, 49.5%) and cluster 3 (average, 55.9%) was evident in the heat map where the sidebar at the top was predominantly green. For CRP, the proportions of participants with high CRP (>2 mg/L) was highest in cluster 1 (large: 32.4%), compared with 13.6% in cluster 2 (thin) and 11.8% in cluster 3 (average). In a multivariate multinomial model that included all 5 predictors, HIV status and CRP remained significant (P < 0.001). Race/ethnicity was also significant (P = 0.03), with a greater odds of being Hispanic in cluster 2 relative to the other 2 clusters. The univariate results were confirmed by the permutation P values. Consistent with the observed heat map, the cluster with the lowest median percent LEGF had the highest proportion of HIV-infected subjects in 96% of 500 bootstrapped samples. Similarly, in 92% of the samples, the cluster with the highest median percent LEGF had the highest proportion of subjects with CRP >2 mg/L.
To investigate whether there were different relationships by HIV status between high levels of CRP (outcome) and cluster, sex, Tanner stage or race/ethnicity (predictors), a logistic regression model was fit for one predictor at a time, separately by HIV status. High CRP was defined as >2 mg/L. In HIV-uninfected subjects, sex and race/ethnicity were not related to high CRP levels after adjusting for cluster. In a model with Tanner stage and cluster, both cluster (P = 0.003) and Tanner stage (P = 0.027) were predictive. Cluster 2 (thin) had no subjects with high CRP (0.0%), cluster 3 (average) had 6.9% with high CRP, and cluster 1 (large) had 31.1% with high CRP. The prevalence of high CRP increased with Tanner stage (9.9% in Tanner 1–2, 22.2% in Tanner 3–4, and 27.3% in Tanner 5). In HIV-infected subjects, only cluster was predictive. As with the HIV-uninfected subjects, cluster 2 (thin) had the lowest percent of subjects with high CRP (15.8%), followed by cluster 3 (average, 17.9%) and cluster 1 (large, 38.5%). Overall prevalence of high CRP was greater in each cluster for the HIV-infected compared with the HIV-uninfected participants.
We have illustrated the use of heat maps to assess relationships among groups of measurements in different domains and to identify participant phenotypes. In this study population, we identified 3 distinct clusters with characteristic phenotypes. One cluster (thin) consisted mostly of HIV-infected participants with abnormal body shape, characterized by lower EXF and dyslipidemia. In contrast, the other 2 clusters (large and average) had a more equal mix of HIV-infected and HIV-uninfected participants, suggesting that host and environmental factors may play more important roles in these phenotypes. The heat maps helped us visualize patterns that support previous observations that some HIV-infected children have a specific phenotype that is probably related to HIV itself or to antiretroviral therapies. However, they also illustrate that with improved health, HIV-infected children and young adults may have a similar risk as HIV uninfected for diabetes and cardiovascular outcomes.
Using the phenotypes identified in the heat maps allowed us to explore more complicated relationships with other participant characteristics and see whether patterns were similar by HIV status. For example, we were able to explore the relationship between phenotype and CRP, a predictor of future cardiovascular disease, and found that the large cluster had a higher proportion of subjects with elevated CRP. The HIV-infected group had higher rates of elevated CRP than the HIV-uninfected in all 3 clusters, suggesting that generalized inflammation is not related only to body phenotype in HIV-infected children and young adults. However, only in the HIV-uninfected subjects did the rates of elevated CRP significantly increase with Tanner stage. These were cross-sectional analyses, so we could not determine temporal relationships between body phenotype and CRP.
Heat map analyses have advantages over standard approaches when investigating differences between groups of subjects. For example, in a dataset with many measurements, variables can be highly correlated (eg, HDL, LDL, total cholesterol), and models may break down. With heat maps, there are no such limitations on the number of variables being used, their correlations, or any assumptions of multivariate normality. Less well-known predictive methods such as random forest or support vector machines offer less insight into associations between variables.
Heat maps are based on cluster methodology, and care is needed in their use. Data must be on a similar scale, different patterns are observed depending on the distance measure selected, and there is subjectivity in choosing the “best” number of clusters. For example, potentially meaningful distinct color patterns were observed within the 3 chosen participant subclusters. It is also important to assess robustness of assignment to clusters. We illustrated approaches using a permutation test and the use of bootstrapping. Despite these caveats, heat maps and multivariate analyses can be used to further explore the relationships between participant clusters and HIV-specific host and environmental factors and could include many more outcomes than are illustrated in this example. Heat maps may prove to be particularly useful for exploratory and hypothesis-generating analyses.
The authors would like to thank the children and young adults who participated in this study, their families, and the entire P1045 protocol team for their contributions and support.