Unsupervised Clustering Reveals Sarcoidosis Phenotypes Marked by a Reduction in Lymphocytes Relate to Increased Inflammatory Activity on 18FDG-PET/CT

Abstract

Introduction: Sarcoidosis is a T-helper cell mediated disease characterized by granulomatous inflammation. We posited that unsupervised clustering of various features in sarcoidosis would establish phenotypes associated with inflammatory activity measured by 18FDG-PET/CT. Our goal was to identify unique features capable of distinguishing clusters and subsequently examine the relationship with FDG avidity to substantiate their potential use as markers for sarcoidosis inflammation. Methods: We performed a retrospective study of a diverse, but primarily African American, cohort of 58 subjects with biopsy proven sarcoidosis followed at the University of Illinois Bernie Mac Sarcoidosis Center and Center for Lung Health who underwent 18FDG-PET/CT scan. Demographic, therapeutic, radiographic, and laboratory data were utilized in unsupervised cluster analysis to identify sarcoidosis phenotypes. The association between clusters, their defining features, and quantitative measurements on 18FDG-PET/CT was determined. The relevance of these features as markers of 18FDG-PET/CT inflammatory activity was also investigated. Results: Clustering determined three distinct phenotypes: (1) a predominantly African American cluster with chronic, quiescent disease, (2) a predominantly African American cluster with elevated conventional inflammatory markers, advanced pulmonary disease and extrathoracic involvement, and (3) a predominantly Caucasian cluster characterized by reduced lymphocyte counts and acute disease. In contrast to the chronic quiescent cluster, Clusters 2 and 3 were defined by significantly greater FDG avidity on 18FDG-PET/CT. Despite similarly increased inflammatory activity on 18FDG-PET/CT, Clusters 2, and 3 differed with regards to extrathoracic FDG avidity and circulating lymphocyte profiles, specifically CD4+ T-cells. Notably, absolute lymphocyte counts and CD4+ T-cell counts were found to predict 18FDG-PET/CT inflammatory activity by receiver operating curve analysis with a 69.2 and 73.42% area under the curve, respectively. Conclusions: Utilizing cluster analysis, three distinct phenotypes of sarcoidosis were identified with significant variation in race, disease chronicity, and serologic markers of inflammation. These phenotypes displayed varying levels of circulating inflammatory cells. Additionally, reduction in lymphocytes, specifically CD4+ T-cells, was significantly related to activity on 18FDG-PET/CT. Though future studies are warranted, these findings suggest that peripheral lymphocyte counts may be considered a determinant of sarcoidosis phenotypes and an indicator of active inflammation on 18FDG-PET/CT.

Keywords: 18FDG-PET/CT; cluster analysis; immunopathogenesis; lymphopenia; phenotype; sarcoidosis.

Figure 1

Box and whisker plots demonstrating baseline laboratory parameters utilized as clustering features of the UIC-Sarcoidosis cohort (n = 58). Values are depicted as median and interquartile ranges for each individual parameter.

Figure 2

Scatterplot visualization of the UIC-Sarcoidosis cohort over 2 dimensions utilizing the t-stochastic neighbor embedding (t-SNE) dimension reduction algorithm. The first and second dimensions are represented by the x-axis and y-axis, respectively. Points on the scatterplot represent individual subjects in the UIC-Sarcoidosis cohort and the distance between points is indicative of dissimilarity between subjects. Colors represent clusters identified by the Modha-Spangler algorithm utilizing partitioning around medoids with Gower's distance for base clustering of mixed data. Overall, Cluster 1 was identified as the largest cluster and comprised 41.38% of the cohort (24/58 subjects). Cluster 2 and 3 represented 25.86% (15/58) and 32.76% (19/58) of the cohort, respectively.

Figure 3

Box and whisker plots demonstrating median and interquartile ranges (kcells/μL) of circulating inflammatory cells in the UIC-Sarcoidosis cohort clusters. Comparisons between clusters with Kruskal-Wallis tests were performed and, if significant (p < 0.05), were followed by post-hoc analysis with Dunn's test and Benjamini-Hochberg (BH) adjustment for multiple comparisons. Kruskal-Wallis test significance is displayed in the corresponding panel and significance of Dunn's test, denoted with an asterisk (*), is displayed above the brackets linking respective box and whisker plots. Specifically, neutrophils (A) were found significantly elevated in Cluster 3 compared to Cluster 1 (BH-adjusted p = 0.0141). Conversely, lymphocytes (B) were found significantly reduced in Cluster 3 compared to Cluster 1 (BH-adjusted p = 0.0207). Neutrophil and lymphocyte counts between Cluster 1 and Cluster 2 and between Cluster 2 and Cluster 3 were comparable. Monocytes (C) did not show any significant differences between clusters. NS, non-significant.

Figure 4

Scatterplot in (A) demonstrates the strong direct relationship (Spearman's correlation coefficient; rho = 0.8329, p < 0.001) that is observed between absolute lymphocyte counts (kcells/μL) and absolute CD4+ T-cells (cells/μL). Box and whisker plots in (B) highlight the differences in absolute CD4+ T-cells (cells/μL) that are observed between clusters. A significant association was identified by Kruskal-Wallis test between clusters and absolute CD4+ T-cells (p = 0.0008). Post-hoc analysis with Dunn's test demonstrated significantly reduced absolute CD4+ T-cell counts, denoted with an asterisk (*), in Cluster 3 compared to Clusters 1 and 2 (BH-adjusted p = 0.0005 and 0.0929, respectively). No difference (NS) in absolute CD4+ T-cell counts was observed between Clusters 1 and 2.

Figure 5

Box and whisker plots in (A) demonstrate the association between UIC-Sarcoidosis cohort clusters and the SUV_Background-to-SUV_Max ratio (SUV ratio). An SUV ratio of 2 was pre-specified as a cutoff value to indicate increased inflammatory activity on 18FDG-PET/CT (PET-positivity) and is depicted in (A) as the dashed black horizontal line. A significant association between clusters and SUV ratio was found by Kruskal-Wallis test (p = 0.0052). Post-hoc analysis with Dunn's test demonstrated significantly elevated inflammatory activity, denoted with an asterisk (*), in Clusters 2 and 3 compared to Cluster 1 (BH-adjusted p = 0.0332 and 0.0136, respectively). No difference (NS) was observed between Cluster 2 and Cluster 3 with regards to inflammatory activity. To assess the probability of PET-positivity based on clusters, a logistic regression model was constructed and the dot plot in (B) highlights the odds ratios derived from the model. Cluster 1 was considered the reference given that it demonstrated significantly decreased 18FDG-PET/CT activity compared to the other two clusters. Ultimately, Cluster 2 and Cluster 3 were found to have a 6 and 6.5-fold greater risk of PET positivity (β = 1.7918 and 1.8718 and p = 0.0132 and 0.0061; respectively).

Figure 6

Receiver operating characteristic (ROC) curves demonstrating the sensitivity and specificity of inflammatory cells as predictors of PET-positivity defined as an SUV ratio > 2. Neutrophils (kcells/μL) demonstrated an area under the curve (AUC) of 45.78% whereas lymphocytes (kcells/μL) and CD4+ T-cells (cells/μL) demonstrated a 69.2 and 73.42% AUC, respectively. No difference was identified between the lymphocyte and CD4+ T-cell AUC utilizing the “bootstrap” method with 1,000 replicates (p = 0.36). Both lymphocytes and CD4+ T-cells demonstrated a significantly greater AUC compared to neutrophils (p = 0.0223 and 0.0029, respectively). Optimal cell count thresholds obtained with “Youden's index” suggest that lymphocyte counts ≤ 1.25 kcells/μL in the UIC-Sarcoidosis cohort are most indicative of PET-positivity with a median sensitivity of 51.72% (95%CI: 34.48–68.97) and a median specificity of 82.76% (95% CI: 68.97–93.1). Similarly, CD4+ T-cell counts ≤ 524.5 cells/μL were most indicative of PET-positivity with a median sensitivity of 68.97% (95%CI: 51.72–82.76) and a median specificity of 72.41% (95% CI: 55.17–89.66).

Figure 7

Box and whisker plots demonstrating log₁₀ normalized median and interquartile ranges of radiomic measurements performed to assess burden of inflammatory activity in the UIC-Sarcoidosis cohort clusters. (A,B) Depict total metabolic volume (TMV) and total lesion glycolysis (TLG). (C–F) Depict intrathoracic metabolic volume MV_IT, intrathoracic lesion glycolysis LG_IT, extrathoracic metabolic volume MV_ET, and extrathoracic lesion glycolysis LG_ET, respectively. Comparisons between clusters were performed with Kruskal-Wallis tests and, if significant (p < 0.05), were followed by post-hoc analysis with Dunn's test. Dunn's test was considered significant if BH-adjusted p < 0.1. Kruskal-Wallis test p-values are displayed in their respective panels; brackets above box and whisker plots denote significance of Dunn's test with an asterisk (*). NS, non-significant. Results of significant Dunn's tests: (A)- Cluster 1 vs. 2 (BH-adjusted p = 0.0237) and Cluster 1 vs. 3 (BH-adjusted p = 0.0410); (B)- Cluster 1 vs. 2 (BH-adjusted p = 0.0170) and Cluster 1 vs. 3 (BH-adjusted p = 0.0293); (E)- Cluster 1 vs. 2 (BH-adjusted p = 0.0096) and Cluster 2 vs. 3 (BH-adjusted p = 0.0409); (F)- Cluster 1 vs. 2 (BH-adjusted p = 0.0058) and Cluster 2 vs. 3 (BH-adjusted p = 0.0381).