Abstract
A practical method to identify people who are most affected by periodontitis in their age group is currently unavailable. We focused on individuals with mean clinical attachment loss (CAL) above the 80th percentile within each of 10 age groups (5-y intervals between 30 and 74 y as well as ≥75 y). We developed predictive models using combined data from 2 cohorts (2009 to 2010 and 2011 to 2012) from the NHANES (National Health and Nutrition Examination Survey; development cohort [DC], n = 6,757), and we carried out external validation using data from a third NHANES cohort (2013 to 2014; validation cohort [VC], n = 3,447). We used 1) age-specific logistic regression models with stepwise selection to identify significant demographic variables, habits, medical conditions, and selected clinical periodontal parameters (proportion of teeth with probing depth ≥4 mm at incisors and molars and with visible [≥2 mm] recession) and to calculate propensity scores (PSs); 2) Youden’s J statistic to select optimum PS cutoffs to maximize diagnostic performance using receiver operating characteristic curves; and 3) bootstrap resampling with 1,000 replicates to validate the age-specific models and adjust the PS and optimal PS cutoffs for overfitting. The bootstrap-adjusted PSs were used as single predictors of mean CAL over the 80th percentile in the VC. The age-specific upper quintiles of mean CAL ranged between 1.63 and 3.24 mm in the DC and between 1.87 and 3.20 mm in the VC. The area under the curve of the models exceeded 0.85 in all age groups in the DC and 0.84 in the VC, indicating well-validated diagnostic performance. In the DC, sensitivity values ranged between 0.75 and 0.97 and exceeded 0.83 in 8 of 10 age groups. Corresponding values in the VC ranged between 0.56 and 0.89 and exceeded 0.68 in 8 of 10 age groups. We conclude that modeling that incorporates readily obtainable variables through a brief patient interview and an abbreviated periodontal examination accurately identifies individuals who are most affected by periodontitis in different ages.
Keywords
Introduction
Global epidemiologic data on the prevalence, extent, and severity of periodontitis in various populations suggest that a substantial proportion is affected and that this population segment increases with age (Kassebaum et al. 2014; Papapanou and Lindhe 2015). Recent advances in our knowledge on the pathobiology of periodontitis have established the inflammatory nature of the disease and its association with dysbiotic biofilms, genetic predispositions, socioeconomic determinants, and environmental exposures (Kinane et al. 2017). Despite variation in the definitions of “severe” periodontitis in the global literature, the data suggest that advanced forms that may result in substantial tooth mortality and loss of function affect approximately 10% to 15% of the population (Papapanou and Lindhe 2015). Although periodontitis of lower severity can still have a negative impact on a person’s oral and overall well-being and may act as a systemic inflammatory stressor that influences general health (Papapanou and Susin 2017), identification of individuals in the most affected segment of the population is important for 1) more accurate stratification that can facilitate the study of the biological determinants of susceptibility and 2) better allocation of therapeutic resources and more effective implementation of secondary prevention strategies. Arguably, the level of severity that signifies high susceptibility and may ultimately jeopardize the dentition should be defined with age-specific, rather than universal, fixed thresholds of attachment loss (Wennström et al. 1990; Lang and Tonetti 2003). For example, incipient disease manifesting itself through a modest amount of attachment loss in a young person may still be a harbinger of advanced disease later in life (Van der Velden et al. 2006).
A recent publication (Billings et al. 2018) described the cumulative distribution of individuals according to mean clinical attachment loss (CAL) in various ages based on data from 2 large population-based epidemiologic studies in the United States (National Health and Nutrition Examination Survey [NHANES] 2009 to 2014) and Germany (SHIP-Trend 2018 to 2012 [Study of Health in Pomerania]). This work defined thresholds of mean CAL that can 1) facilitate a comparative assessment of periodontitis severity, 2) identify where a person ranks with respect to a “severity scale,” and 3) be used to titrate the frequency or intensity of preventive/therapeutic interventions. However, a prerequisite for utilizing the published thresholds is availability of full-mouth CAL measurements that, contrary to their common use in research settings, are seldom carried out in the general dental practice.
The aim of this work was to develop a system that utilizes easily retrievable information through a patient interview and an abbreviated clinical periodontal examination that does not require assessment of CAL and can be used to identify individuals above the 80% percentile of periodontitis severity for their age. We developed this system using available epidemiologic data representative of the US adult population, and we examined its external validity in an independent, population-representative sample of US adults derived at a later time point.
Materials and Methods
The analyses were conducted in accordance with TRIPOD (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis; Collins et al. 2015).
Source of Data and Study Participants
Data were derived from 3 cohorts of the NHANES (2009-2010, 2011-2012, and 2013-2014), which uses a stratified multistage probability cluster sampling design of civilian noninstitutionalized subjects who reside in households in the conterminous United States.
Study Design, Sample Size, and Handling of Missing Data
The total number of individuals ≥30 y old in the 3 selected NHANES cohorts was 14,556, of whom 11,753 were periodontally examined and 10,677 had valid periodontal data and were eligible for analyses. This number was further reduced to 10,204 after deletion of 473 participants (4.4% of those with valid periodontal data) because of missing/incomplete data for selected variables used in our analyses. We combined the 2009-2010 and the 2011-2012 NHANES cohorts to create a single cohort to be used in the development and internal validation of the predictive models (development cohort [DC], n = 6,757). The third NHANES cohort (2013-2014) was used to externally validate the predictive models (validation cohort [VC], n = 3,447].
Data Stratification
To account for differences in periodontitis extent and severity and related risk factors with age (Billings et al. 2018), persons in the DC and VC were stratified into 9 age groups with 5-y intervals between 30 and 74 y and a 10th group ≥75 y old. The mean sample size per age group was 510 (SD = 120) with an approximate 1:1 male:female ratio.
Main Outcome
CAL measurements recorded at 6 sites per tooth at all teeth excluding third molars were averaged to calculate a mean CAL for each individual. Persons belonging in the upper quintile of mean CAL in each age group, representing those most affected for their age, were identified separately in the DC and VC. Membership in this group was used as the main dichotomous outcome in all analyses.
Potential Predictors
Potential predictors of the main outcome were selected among variables with established/tentative association with periodontitis that 1) were available in all 3 NHANES waves, 2) were deemed to be easily retrievable in the setting of a dental examination, and 3) would not reduce the eligible sample by >10% because of missing data. These included 1) demographic variables (sex, race/ethnicity, educational level, and insurance status), 2) medical conditions (diabetes, arthritis, hypertension, asthma, cancer, and thyroid disorders), and 3) tobacco consumption (see Appendix for exact variable definitions). In addition, 3 variables readily assessed in a clinical periodontal examination were also tested: proportion of teeth present (excluding third molars) with visible (≥2 mm) gingival recession on any surface and proportion of teeth with ≥4- and ≥5-mm probing depth among incisors and molars present (excluding third molars).
Statistical Analysis
Predictive models were developed in the DC as follows.
Separately in each age group, multiple logistic regression analysis was performed to model the probability of a participant to have a mean CAL above the 80th percentile. All 14 potential predictors, including their 2-way interactions, were considered in the stepwise selection procedure with an entry probability of 0.05 and a stay probability of 0.05. The Appendix describes the frequency distribution of the finally retained predictors in the DC and VC.
For each person and with use of XBn, the fitted value based on the retained predictors in each age-specific model, a propensity score (PS) was computed as follows: PS = exp(XBn) / [1 + exp(XBn)]. Next, a simple logistic regression with the PS as a single predictor of the dichotomous outcome was carried out, and the area under the curve (AUC) was calculated. Optimal PS cutoff thresholds were determined according to the Youden’s J statistic (sensitivity + specificity − 1; Youden 1950; Fluss et al. 2005). Bootstrap internal validation with 1,000 resampling replicates (Harrell 2015) was performed for each age-specific model. Bias-adjusting intercepts and slopes created an adjusted PS for each participant to further correct for model overfitting as follows:
Bootstrap-adjusted optimal PS cutoff thresholds (Table 3, last column) were obtained by replacing the “PS” in the formula with the optimal PS cutoff values (Table 3, fourth column).
External validation was conducted in the VC with the age-specific bootstrap-adjusted PS to predict mean CAL above the 80th percentile in a simple logistic regression model, and the AUC was calculated. Next, in each age group, the adjusted PS was dichotomized with the adjusted PS cutoff threshold (Table 3, last column) and cross-tabulated with the dichotomized outcome to form a 2-way contingency table, from which sensitivity, specificity, positive predictive value, negative predictive value, and accuracy were calculated.
SAS (version 9.4; SAS Institute) was used to run the stepwise logistic regression. Package “rms” of the R Project for Statistical Computing (version 3.5.1) was used to conduct the bootstrap internal validation. All other analyses were conducted with SPSS (version 25.0; IBM).
Target Use and Users
The intended use of the developed predictive models is to assist clinicians in the identification of patients who have developed CAL of a severity that places them in the upper quintile for their age.
Results
Outcome Distribution
The mean CAL ranged between 1.24 mm (30 to 34 y) and 2.38 mm (65 to 69 y) in the DC and between 1.47 mm and 2.22 mm in the VC. The 80th percentile of CAL ranged between 1.63 mm (30 to 34 y) and 3.24 mm (65 to 69 y), while corresponding values were between 1.87 and 3.20 mm in the VC (Table 1).
Clinical Attachment Loss by Age Group in the Development and Validation Cohorts.
Values are presented in millimeters.
NHANES, National Health and Nutrition Examination Survey.
Identified Predictors
Two continuous periodontal variables—proportion of teeth (excluding third molars) with visible recession (≥2 mm) and proportion of teeth with PD ≥4 mm among incisors and molars (excluding third molars)—were identified as significant predictors (P < 0.05) across all age groups (Table 2). Nine categorical variables (sex, race/ethnicity, educational level, insurance status, current smoking, diabetes mellitus, thyroid disorder, arthritis, and hypertension) emerged as significant predictors in some, but not all, age groups. Significant interaction terms were included as indicated.
Variables Retained in the Predictive Model by Age Group.
PD, probing depth.
Third molars excluded.
Third molars excluded.
Interaction terms were included for variables sharing the same superscript letter.
Performance Measures of the Fitted and Bootstrap-Corrected Models
Table 3 presents the diagnostic performance of the fitted and bootstrap-corrected age-specific models. The AUC exceeded 0.85 in all fitted models and ranged from 0.855 (30 to 34 y) to 0.986 (55 to 59 y). The maximum Youden’s J score ranged from 0.586 (with 78.2% sensitivity and 80.4% specificity) in the 30 to 34 y old group to 0.891 (with 95.2% sensitivity and 94.0% specificity) in the 60 to 64 y old group.
Diagnostic Performance of the Fitted and the Bootstrap-Corrected Models.
AUC, area under the curve; PS, propensity score; PV+, positive predictive value; PV–, negative predictive value; Youden’s J, sensitivity + specificity − 1.
Bootstrap-corrected optimal PS cutoff value.
Internal validation with bootstrap correction resulted in bias-adjusting slopes that remained close to 1 in all age groups, ranging from 0.9200 (60 to 64 y) to 0.9874 (50 to 54 y), while the intercept remained close to 0 in all groups, ranging from −0.0007 to −0.0594. The AUC values of the bootstrap-fitted model remained similar to that of the final model (data not shown). There was a slight overall increase in the optimized PSs after bootstrap correction.
External Validation
Use of the age-specific bootstrap-corrected PSs in the validation cohort to identify membership in the upper quintile of mean CAL per age group resulted in AUC values ranging from 84.2% (30 to 34 y) to 98.3% (55 to 59 y; Table 4). The lowest sensitivity was observed in the 4 youngest groups (55.7% to 69.3%) but was >70% in all other groups. Specificity values exceeded 90% in all groups but the youngest (83.0%), and accuracy exceeded 80% in all and 90% in 7 of the 10 groups, indicating that the external validation was satisfactory.
Age-Specific Model Performance Summary: Validation Cohort.
AUC, area under the curve; PV+, positive predictive value; PV–, negative predictive value; Youden’s J, sensitivity + specificity – 1.
(Number of true positive subjects + number of true negative subjects) / total number of subjects.
Discussion
In this study, we sought to generate algorithms that can reliably identify individuals whose periodontitis severity ranks them above the 80th percentile of mean CAL among their peers of corresponding age. To that end, we first used data from a nationally representative population-based sample in the United States to develop age-specific models based on variables that are easily retrievable through an interview and an abbreviated clinical periodontal examination. After the initial identification of predictors of the defined outcome (CAL exceeding the 80th percentile in the particular age), we calibrated the models using bootstrap resampling to correct for bias due to overfitting. In an ultimate step, we tested the external validity of the models using representative samples of the US population, drawn at a later time point, in corresponding ages. Overall, the diagnostic performance of the developed models in the independent cohort was highly satisfactory, with AUC exceeding 0.84 in all age groups and with sensitivity values exceeding 0.70 in 6 of 10 age groups.
Several diagnostic or predictive models related to periodontitis have been published in the literature over the past 2 decades. Most of these efforts have attempted to predict periodontitis progression, broadly defined as longitudinal loss of clinical attachment or alveolar bone and/or incident tooth loss (Page et al. 2002; Trombelli et al. 2009; Lindskog et al. 2010; Morelli et al. 2017; Martinez-Canut et al. 2018; Morelli et al. 2018), with a number of publications placing focus on events occurring during the course of supportive periodontal therapy (Lang and Tonetti 2003; Trombelli et al. 2017). However, in a recent systematic review (Du et al. 2018) that critically assessed the performance of prediction models for the incidence and progression of periodontitis, studies were generally characterized by a poor level of reporting, despite an alleged “good” ability to discriminate among people at risk for periodontitis. Notably, only a subset of the reviewed studies carried out external validation of the models in independent cohorts (Page et al. 2002; Morelli et al. 2017; Martinez-Canut et al. 2018; Morelli et al. 2018).
The approach adopted in our work was rather different: instead of attempting to predict risk for periodontitis progression, we focused on the identification of individuals whose clinical presentation suggests that they are “outliers” with respect to periodontitis severity when compared with their peers of similar age. We set the threshold at the upper quintile, as earlier introduced by Locker and Leake (1993). The upper quintile has been widely used in the medical and dental literature to define the most affected portion of the population (Brameld and Holman 2005) and particularly in studies of the social determinants of health (Borrell and Crawford 2012). Given that most global estimates of severe periodontitis in various populations range between 11% and 15% (Kassebaum et al. 2014; Papapanou and Lindhe 2015), we focused on the upper quintile to identify the fraction of the population that is most periodontitis affected. However, our approach was age-specific because a seemingly innocuous amount of attachment loss of a couple of millimeters may signify high susceptibility to periodontitis in young people but be rather inconsequential in an elderly person. Using an analogy from the contemporary diabetes metabolic control guidelines (American Diabetes Association 2019), stricter treatment targets are adopted in young patients with diabetes mellitus, while values well over 7% hemoglobin A1C may be acceptable treatment targets in older patients. Our work is therefore purposefully based on age-specific quintiles of mean CAL rather than on fixed, arbitrary thresholds of CAL that are supposed to represent severe disease across the age spectrum.
The value of successful identification of individuals in the upper quintile is obvious: accurate detection of this highly susceptible group that has experienced a disproportionate loss of periodontal tissue support would be useful for decision making and optimization of resource allocation. Indeed, those in the upper quintile would conceivably benefit from more intense therapeutic and preventive measures, including aggressive control of risk factors and more frequent maintenance visits.
A handful of studies in the literature have attempted to develop algorithms to cross-sectionally identify subjects exceeding a certain level of periodontitis severity. For example, Leite et al. (2017) used decision tree analyses incorporating clinical, sociodemographic, and general health information to classify people of fairly young age (24 to 31 y) according to different periodontitis definitions and reported AUC values ranging from 0.530 to 0.670, depending on the thresholds used. These values are considerably lower than the AUC observed in the youngest cohort of the present study (30 to 34 y; 0.855 and 0.842 in the DC and VC, respectively). Importantly, these authors noted that deep periodontal pockets are not necessarily a prominent feature of periodontitis in young ages and that use of additional variables is necessary to identify affected patients. Recently, Montero et al. (2019) used NHANES 2011 to 2012 data to develop a predictive model to identify people aged >30 y with moderate or severe periodontitis, according to the definition of the Centers for Disease Control and Prevention / American Academy of Periodontology (Eke et al. 2012). In that work, a predictive model that incorporated age, sex, ethnicity, hemoglobin A1C levels, and smoking habits detected moderate/severe periodontitis with 70% sensitivity and 67.6% specificity. However, the model used a fixed definition of periodontitis across the entire age spectrum, and no data of age-specific diagnostic performance were presented. In addition, the authors carried out correction for overfitting of their model using bootstrap resampling, without external validation.
In contrast, in our study 1) we used age-dependent definitions of advanced mean CAL, 2) developed models that incorporated age-specific predictors in different ages, and 3) carried out internal and external validation of the developed models. As noted in earlier publications, a meaningful definition of disease (Scadding 1996) and severe periodontitis in particular (Papapanou 2012) can be derived only after taking into account what is the “norm” for those in a relevant reference group (i.e., peers of similar age). An age-dependent assessment of periodontitis severity, explicitly recognized as an important biological feature of the disease and partly reflected in the grade vector of the new classification system for periodontitis (Papapanou et al. 2018; Tonetti et al. 2018), is meaningful as 1) some level of attachment loss is increasingly common with age and 2) incipient attachment loss in young age is an early sign of disease susceptibility. Based on observations in 2 nationally representative populations in the United States and Europe (Billings et al. 2018) indicating that a substantial component of severe attachment loss in all age groups is manifested through gingival recession, we selected to include the proportion of teeth with visible (≥2 mm) recession along with the proportion of teeth among molars and incisors with probing depth ≥4 mm as potential predictors in our models, and both variables were retained as statistically significant in all age groups. We also tested multiple additional variables, and sex, race/ethnicity, education level, insurance status, and current smoking, as well as the presence of diabetes mellitus, thyroid disorder, arthritis, and hypertension, were ultimately retained as statistically significant predictors in certain, but not all, age groups (Table 2). Thus, all parameters included in our predictive models are easily retrievable in the clinical setting of a general dental practice through a combination of a simple patient interview, a visual inspection of teeth with recession, and probing assessment of pocket depth limited to index teeth. Last, the validation of our models in an independent sample and their overall satisfactory diagnostic performance indicate that the suggested algorithms are sufficiently robust in detecting with high accuracy people who are above the 80th percentile of periodontitis severity in their age in the United States.
A number of limitations of our study must be acknowledged. First, we did not have access to data from individuals aged <30 y; therefore, we were unable to develop predictive models for advanced CAL of early onset. Future studies need to address this gap, given the importance of early detection of young, susceptible people before the development of substantial, irreversible tissue support loss. In addition, the lowest sensitivity of our models in the VC was noted in ages 40 to 44 y, suggesting that a sizable portion of people in the upper quintile of attachment loss may still remain undetected in these ages. Last, the external validation of our models was carried out in an independent cohort from the same source population, albeit one recruited at a later time point and significantly different from the DC with respect to several background characteristics and periodontal status (Appendix). Consequently, additional adjustments to the algorithms will be required when applying our models in other world regions with substantially different distribution/severity of periodontitis than in the United States. In the Appendix, we provide step-by-step instructions for outcome and variable definitions, calculation of adjusted PSs, and statistical code to be used when applying our predictive models in other source populations outside the United States.
In conclusion, the proposed approach is a simple way by which members of the oral health care team can identify those mostly affected by periodontitis across the age spectrum in the United States. Although the most accurate means of identifying these individuals remains the assessment of full-mouth CAL, our methodology offers a less labor-intensive alternative. We welcome further examination of the robustness of our predictive models when applied to other populations.
Author Contributions
J.A. Shariff, contributed to data acquisition and analysis, drafted the manuscript; B. Cheng, contributed to design, data analysis, and interpretation, critically revised the manuscript; P.N. Papapanou, contributed to conception, design, and data interpretation, drafted and critically revised the manuscript. All authors gave final approval and agree to be accountable for all aspects of the work.
Supplemental Material
DS_10.1177_0022034519884518 – Supplemental material for Age-Specific Predictive Models of the Upper Quintile of Periodontal Attachment Loss
Supplemental material, DS_10.1177_0022034519884518 for Age-Specific Predictive Models of the Upper Quintile of Periodontal Attachment Loss by J.A. Shariff, B. Cheng and P.N. Papapanou in Journal of Dental Research
Footnotes
Acknowledgements
The authors thank Drs. Bruce Dye and Monisha Billings for helpful discussions at the planning stages of this work.
A supplemental appendix to this article is available online.
This work was supported by the Division of Periodontics, Section of Oral, Diagnostic and Rehabilitations Sciences, Columbia University.
The authors declare no potential conflicts of interest with respect to the authorship and/or publication of this article.
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
