Systemic Multimorbidity Clusters in People with Periodontitis · Larvin H, Kang J, Aggarwal VR et al.

Methods

This cross-sectional study is based on participants from the NHANES between 2009 and 2014.

Database

Data from the NHANES were utilized in this analysis. The NHANES program is conducted over 2 y by the Centers for Disease Control and Prevention. The data set contains self-reported survey responses to questions related to demographics (age, sex, and ethnicity), socioeconomics (household income quintile), and health (smoking status, health conditions). Oral health information from clinical dental examinations and physiologic measurements such as body mass index and blood pressure from medical examinations by trained health professionals are also captured. Each cycle contains approximately 10,000 participants and is representative of the United States across all age groups.

Study Sample

For this analysis, we used NHANES cycles between 2009 and 2014 that comprise information from periodontal examinations. Participants were eligible for periodontal examination if they were aged ≥30 y. Participants who were edentulous and those who had a heart transplant, artificial heart valve, or congenital heart disease were excluded, as were those who required antibiotic prophylaxis prior to examination. To explore multimorbidity, we further excluded participants who did not have a systemic disease diagnosis (Appendix Fig. 1).

Periodontitis Case Definition

The periodontal examination was conducted at a mobile center by trained dentists who measured clinical attachment loss (CAL) and probing depth across 6 interproximal sites per tooth, excluding third molars. Data were inputted by a health technician. The interclass correlation coefficients of mean attachment loss were ≥0.80 for the included cycles and were in the main more reliable than probing depth measurements. We used the proportion of sites with CAL ≥ 3 mm as a measure of extent of periodontitis, which we split into 4 equal quartiles (Wright et al. 2020). We allocated each participant into 1 of the 4 quartiles according to one’s CAL measures. Quartile 1 represents people with mild/less severe periodontitis, while quartile 4 represents people with more severe periodontitis.

Systemic Disease

A range of systemic diseases was determined by collating participant responses to the question “Have you even been told by a doctor or health professional that you have . . . ?” followed by 1 of the following diseases: diabetes, cancer, heart attack, coronary heart disease, congestive heart failure, angina and stroke, hypertension, arthritis, liver disease, osteoporosis, thyroid disorder, emphysema, and bronchitis. Obesity was defined by body mass index ≥30 kg/m2 (National Institutes of Health 1998). Data on osteoporosis were collected only in NHANES cycles 2009 to 2010 and 2013 to 2014.

Statistical Analysis

Participant characteristics are presented as mean and standard deviation for continuous variables and frequency and percentage for categorical variables. We reported the overall characteristics for the study population and in people across CAL quartiles. As multimorbidity is highly associated with age and sex, we performed 1:1 matching for age and sex across CAL quartiles to ensure that any systemic multimorbidity clusters were not attributed to the difference in age and sex.

Hypergraphs are based on an artificial intelligence output of network analysis to illustrate and explore hyperedges and the hypernodes within them. In a hypergraph, 1 edge can join an infinite number of nodes at once, as opposed to traditional graphs whereby an edge connects only 2 nodes. This technique has demonstrated its utility for illustrating multimorbidity clusters in large-scale data such as electronic health records (Rafferty et al. 2021). Nodes (N denotes number of nodes) in a hypergraph represent single diseases, and edges (E denotes number of edges) represent a disease set (multimorbidity cluster; Fig. 1). An incidence matrix (M) with dimension N × E represents an unweighted hypergraph. The adjacency matrix (A) can be derived from an incidence matrix to allow for further metrics, such as centrality. The adjacency matrix A can be calculated as A = M T M – DN, where T denotes the transpose and DN is the degree to which the node is connected to any number of edges, with zeros on the diagonal representing that a node cannot be connected to itself.

Figure. Common systemic multimorbidity clusters identified through hypergraph analysis in people with periodontitis. Nodes represent diseases, and hyperedges (shaded areas) represent multimorbidity clusters. For illustrative purposes only; not all systemic multimorbidity clusters are shown. CHF, congestive heart failure; CHD, coronary heart disease. (See the original article for the figure.)

To ensure accurate representation of the study sample, the edges of hypergraphs were weighted by WE, which represents a measure of the number of people who have the disease conditions that are connected to the edge. To account for overlapping sets, in this study the overlap weighting coefficient (WE) was calculated by the number of people with all the diseases in the set divided by the minimum number of people with 1 of the diseases. Adding weights to the hypergraph results in the following formula for the adjacency matrix: A = M T WEM – DN.

The dual hypergraph is symmetrical to a hypergraph; however, the dual hypergraph allows for centrality of multimorbidity clusters to be measured—that is, the degree to which nodes are connected to each multimorbidity cluster. The adjacency matrix (A*) of the dual hypergraph is weighted by the overlap weighting coefficient (WE) and the number within the cohort with a disease (WN, or crude prevalence): A*=WE(MWNMT−DE)WE.

The hypergraph can be analyzed via centrality metrics. In accordance with graph theory, the adjacency matrix of the hypergraph represents the weighted strength of connections within the hypergraph (Rafferty et al. 2021). To compute the eigenvector centrality, we calculate the eigenvectors of the adjacency matrix A. Eigenvector centrality of the adjacency matrix A refers to the degree in which a given node (disease) is connected to other nodes with their own levels of connectivity. For example, a higher centrality value suggests that a disease cluster is more common than other disease clusters with lower centrality. We applied centrality metrics (eigenvector) and a dual hypergraph (whereby the nodes and edges in the hypergraph are alternated) to the adjacency matrix to determine the most central disease and systemic multimorbidity clusters across CAL quartiles. In addition to eigenvector centrality, we plotted heat maps to illustrate the co-occurrences of 2 single diseases (WN) and bar plots to show the highest weighted systemic multimorbidity clusters (WE).

Further to the analysis as applied to the whole study population, we stratified the data by ethnicity, smoking status, and household income to show the differences to single-disease centrality of systemic multimorbidity clusters in these subgroups. We also examined the effect of missing survey responses and osteoporosis information (NHANES cycle 2011 to 2012) by sensitivity analysis in which we compared the results with this cycle removed. Data management, analysis, and hypergraph analysis were conducted in R version (4.0.0; R Core Team 2017). This report conforms to STROBE guidelines.