Systematic review of studies examining contribution of oral health variables to risk prediction models for undiagnosed Type 2 diabetes and prediabetes

Abstract Objective To conduct systematic review applying “preferred reporting items for systematic reviews and meta‐analyses statement” and “prediction model risk of assessment bias tool” to studies examining the performance of predictive models incorporating oral health‐related variables as candidate predictors for projecting undiagnosed diabetes mellitus (Type 2)/prediabetes risk. Materials and Methods Literature searches undertaken in PubMed, Web of Science, and Gray literature identified eligible studies published between January 1, 1980 and July 31, 2018. Systematically reviewed studies met inclusion criteria if studies applied multivariable regression modeling or informatics approaches to risk prediction for undiagnosed diabetes/prediabetes, and included dental/oral health‐related variables modeled either independently, or in combination with other risk variables. Results Eligibility for systematic review was determined for seven of the 71 studies screened. Nineteen dental/oral health‐related variables were examined across studies. “Periodontal pocket depth” and/or “missing teeth” were oral health variables consistently retained as predictive variables in models across all systematically reviewed studies. Strong performance metrics were reported for derived models by all systematically reviewed studies. The predictive power of independently modeled oral health variables was marginally amplified when modeled with point‐of‐care biological glycemic measures in dental settings. Meta‐analysis was precluded due to high inter‐study variability in study design and population diversity. Conclusions Predictive modeling consistently supported “periodontal measures” and “missing teeth” as candidate variables for predicting undiagnosed diabetes/prediabetes. Validation of predictive risk modeling for undiagnosed diabetes/prediabetes across diverse populations will test the feasibility of translating such models into clinical practice settings as noninvasive screening tools for identifying at‐risk individuals following demonstration of model validity within the defined population.


| INTRODUCTION
More than 30 million adults (over 9% of the US population) have diabetes, and nearly one in three (approximately 84 million individuals) are prediabetic. The 2020 national statistics report on diabetic disease burden in the United States estimated that 10.5% of the population has diabetes with 21.4% of individuals with diabetes unaware of their status, while 34.5% of the population has prediabetes (Centers for Disease Control and Prevention, 2020). Like diabetes, periodontitis has also achieved epidemic status in the United States based on recently updated estimates of periodontitis prevalence projected by Eke et al. (2000) following examination of the National Health and Nutrition Examination Survey (NHANES) population data. Applying three case definitions, the authors defined a prevalence rate of 42%, a sixfold increase over diabetes prevalence. Among those with periodontitis, 7.8% met definitions aligning with severe periodontitis (Eke et al., 2000). A growing evidence base reinforced by both systematic review (Borgnakke et al., 2013) and meta-analysis (Corbella et al., 2013;Wang et al., 2014) provides evidence of an adverse impact of periodontitis on glycemic control. The escalating prevalence of undiagnosed Type 2 diabetes mellitus and prediabetes (T2DM/prediabetes) and periodontitis in the global population are noteworthy in light of these biological interactions.
Bidirectional interaction and exacerbation between T2DM and periodontitis has been proposed (Casanova et al., 2014;Llambés et al., 2015;Sgolastra et al., 2013). Interaction between periodontitis and T2DM is supported by multiple reports of heightened periodontitis severity in association with uncontrolled T2DM and improved glycemic control following periodontitis treatment (Borgnakke et al., 2013;Casanova et al., 2014;Corbella et al., 2013;Demmer et al., 2010;Eke et al., 2000;Glurich et al., 2013;Llambés et al., 2015;Sgolastra et al., 2013;Wang et al., 2014). A study that conducted longitudinal monitoring of normoglycemic individuals with and without periodontitis noted an incremental fivefold increase in HbA1c levels overtime only in individuals with periodontitis (Demmer et al., 2010). This observation suggests a potential role for periodontitis in the pathogenesis of T2DM.
Underlying inflammatory processes are posited to contribute to the mutual exacerbation of T2DM and periodontitis (Glurich et al., 2013).
Clinical challenges are faced by both medical and dental providers in managing patients with periodontitis and T2DM/prediabetes. Chronic local and systemic inflammation impedes efforts by medical providers to control hyperglycemia in diabetic patients with comorbid periodontitis, while dentists face challenges in preventing or resolving periodontitis in patients with uncontrolled blood sugar levels.
Importantly, both periodontitis and T2DM represent potentially modifiable conditions if detected and managed prior to becoming chronically established. Ideal clinical management strategies include an integrated medical-dental care delivery approach that targets patient education and activation, lifestyle and behavioral interventions, early clinical intervention, and longitudinal monitoring for disease recurrence or progression (Glurich et al., 2017;Glurich, Schwei, et al., 2018;Shimpi et al., 2016;Shimpi et al., 2020). The Center for Disease Control and Prevention's National Prevention Program estimates that such interventions can lower the risk of developing T2DM by nearly 60% (70% after age 60 years) (Centers for Disease Control and Prevention, 2018). However, traditional silo-ed medical and dental healthcare delivery models in the United States pose substantial barriers to holistic care delivery for the management of cumulative health risks associated with these conditions. Following the review of 14 studies that collectively enrolled >32,000 patients, a 2018 consensus report by the International Diabetes Federation and European Federation of Periodontology concluded that a solid evidence base supported links between periodontitis and T2DM and its complications (Sanz et al., 2018).
Moreover, a recent systematic review of 10 field trials exploring chairside screening of undiagnosed T2DM/prediabetes in the dental setting consistently demonstrated high levels of prediabetes across diverse dental primary care practices (Glurich, Bartkowiak, et al., 2018). These findings collectively promote the adoption of an alternative integrated health care delivery paradigm with focus on early detection and management of undiagnosed T2DM/prediabetes.
Such an integrated care model would include cross-disciplinary screening in both medical and dental settings with appropriate triage across the medical and dental domains.
Assessing patient risk for a diabetes involves noninvasive screening of patients by assessing their status with respect to informative candidate variables. Patient status for defined risk variables is assessed by risk prediction modeling applying statistical approaches including multiple regression modeling. Analysis of data available in medical electronic health records (EHR) that predict relative risk for diabetes include family history, clinical, pharmaceutical, demographic, and environmental factors. Data modeling holds promise for noninvasive detection of undiagnosed disease based on information already available in the EHR.
Historically, over 80 publications have reported on the evaluation of the performance of diabetes risk prediction models in various populations. However, few models have included oral health variables.
Over the past 10 years as the evidence base suggesting a bidirectional association between T2DM and periodontitis progression increased, researchers began exploring the application of diabetes risk modeling in dental settings as a noninvasive approach to risk assessment. Such studies also tested risk prediction model performance incorporating dental variables and compared performance of these models to biological determination of glycemic levels as the gold standard (Borrell et al., 2007;Holm et al., 2016;Lalla et al., 2011Lalla et al., , 2013Li et al., 2011;Strauss et al., 2010).
A systematic review conducted by Collins et al. (2011) evaluated methodology used in studies that developed multivariable diabetes risk prediction models from 1980 to 2011, and included studies that had evaluated oral health factors. While search terms used by Collins et al. (2011) and our study varied somewhat, both search strategies identified the same articles within the temporal window of their study that had applied multivariable analysis. The current study included those studies also reviewed by Collins et al. (2011) that had applied rigorous modeling approaches and also met the eligibility criteria of the current study. However, the temporal frame of the current study was extended to include eligible studies published through July 31, 2018. The rationale for inclusion of later studies was to identify additional models that met inclusion criteria of the current study with the same stringency criteria applied by Collins et al. (2011) and further explored inclusion of oral health variables in predictive models examining risk for undiagnosed Type II diabetes/prediabetes. The systematic review was conducted on eligible studies that applied multivariable regression modeling or bioinformatics approaches with a requirement for inclusion of oral health/dental variables.

| Systematic review
The current study met exemption criteria of the Institutional Review Board in those research activities were limited only to the review of historically published literature and included no research activities involving human subjects or animals. The systematic review was conducted and reported in accordance with "preferred reporting items for systematic reviews and meta-analyses" (PRISMA) statement (Moher et al., 2009). The research question in the current study was defined within the patients-(intervention/exposure/prognostic factor)-comparison group-outcome) (PICO) framework as follows: "P" = persons at risk for undiagnosed T2DM or prediabetes; "I" = prognostic factor: oral health variable(s) alone or in combination with other risk variables; "C" = comparison group: comparisons with a nondiabetic population; and "O" = outcome: predictive capacity defined by the derived predictive model.

| Literature identification selection and review
Articles published from January 1, 1980 through July 31, 2018 including those meeting statistical rigor previously defined in the systematic review by Collins et al. (2011) were identified by systematic searches conducted in PubMed, Web of Science databases, and Gray Literature with last search conducted in August, 2018. Literature was retrieved using the search strategy defined in Figure 1. Literature searches were conducted by a medical librarian with extensive expertise and experience in the conduct of systematic reviews.
Articles qualifying for inclusion were required to meet the following eligibility criteria.

| Inclusion criteria
Original articles were required to: 1. specify the creation of T2DM/prediabetes risk prediction models in undiagnosed individuals by modeling oral/dental and other candidate risk variables applying multivariable regression modeling or biomedical informatics approaches such as machine learning or classification and regression trees (CART) analysis; 2. include oral/dental health-related variables; 3. be published between January 1, 1980 and July 31, 2018; and 4. report on model performance relative to predicting risk for T2DM/prediabetes. Articles that compared the performance of models that included oral and other candidate risk factors with and without biological screening outcomes were also retained and systematically reviewed.

| Exclusion criteria
Articles were excluded if they were a publication type other than an original article, predicted risk for other diabetes types, fell outside of the defined date range, were in a language other than English, did not include at least one oral health-related variable, and did not evaluate predictive capacity of the models.
2.3 | Template development, data collection, interrater reliability, and quality assessment A review protocol ( Figure 2) was created to ensure systematic abstraction and data capture across each manuscript. The template was pilot tested by three reviewers to ensure cohesive understanding of the definition of all abstraction terminology among reviewers. Title and abstract review were first conducted to identify articles requiring further review.
Following the initial review of n = 8 articles by two reviewers (N.S. and I.G.), inter-rater assessment as done to evaluate the interpretation of defined abstraction terms and inform finalization of the research electronic data capture (REDCap) (Harris et al., 2009) template used for fulltext review of remaining articles. A second inter-rater assessment of an additional three articles tested agreement between data collected via the REDCap-based template used by each reviewer. Quality assessment was further conducted following abstraction of the same articles by a third reviewer (G.J.) to replicate good inter-rater agreement using the template. The final inter-rater assessment was conducted to confirm high inter-rater agreement (n = 7) between the reviewers (N.S. and I.G.), and was reported as the percentage agreement between reviewers.
Summary measures assessed included outcomes of predictive modeling.
The most frequent analyses presented included area under the curve (AUC), and/or assessment of sensitivity, specificity, and positive and negative predictive values.

| Bias assessment
The risk of bias was assessed by two reviewers (I.G. and R.B.). The prediction model risk of assessment bias tool (PROBAST) (Wolff et al., 2019) was applied. The PROBAST tool was published in 2019 specifically to address the absence of a bias assessment tool for studies developing or validating predictive models.

| OUTCOMES OF SYSTEMATIC REVIEW
Search terms and outcomes of the systematic search strategy summarized in the PRISMA flow chart are shown in Figure 1.
Among 141 articles retrieved, 95 articles defined candidate variables for predicting risk for undiagnosed T2DM/prediabetes, and 71 studies met eligibility for full review. Among the 71 articles, 10 studies were identified that included oral health or dental F I G U R E 1 Search terms and PRISMA flow diagram. This figure summarizes the search terms used to identify potentially eligible publications and the outcome of the PRISMA review process denoting numbers of publications initially identified and screened for eligibility and the final number of publications meeting criteria for full systematic review. PRISMA, preferred reporting items for systematic reviews and meta-analyses (Moher et al., 2009) variables. However, three were found to be review articles or commentaries and were excluded. Thus, seven publications met the criteria and were systematic reviewed. Inter-rater agreement across data elements abstracted during final assessment achieved 99%. Table 1 summarizes an overview of the outcomes of the systematic review. Collectively, the studies demonstrated that oral variables contribute to models that predict the presence of undiagnosed T2DM/prediabetes identified by biological testing, both independently and in combination with other candidate risk factors. At the conclusion of systematic review, it was determined that further metaanalysis was not feasible based on detection of high variability across study design of the seven reviewed studies, goal and outcomes, study population characteristics, and settings in which these studies were Results of bias assessment applying the PROBAST are shown in Table 2. Potential sources of bias associated with most studies included the use of self-reported data and incomplete reporting of analytical detail. The risk of bias or applicability consistently ranged from "low" to "unclear" across studies systematically reviewed.

| DISCUSSION
The current study undertook a systematic review of the application of predictive modeling for detection of undiagnosed T2DM/prediabetes applying oral health variables as candidate predictors. This study builds upon an earlier study by Collins et al. (2011), which had systematically examined the methodological rigor of studies published through 2010 whose goal was the prediction of undiagnosed T2DM/ prediabetes. Guided by the outcomes of their analysis, only studies that applied multivariable regression modeling to define predictive variables were considered for inclusion.
Most multivariable regression modeling across studies systematically reviewed was generally undertaken as "proof of concept" to test F I G U R E 2 Flow chart of data abstraction and systematic review protocol. (a) Provides an overview of the study screening protocol including the determination of inter-rater reliability. (b) Summarizes the abstraction protocol including publication types screened, screening of methodological approach and variables assessed by the publications, and conduct of the systematic review and bias evaluation applying PRISMA and PROBAST on articles meeting eligibility criteria. PRISMA, preferred reporting items for systematic reviews and meta-analyses; PROBAST, prediction model risk of assessment bias tool Whereas all of the studies systematically reviewed herein focused on assessing the prediction of undiagnosed T2DM/prediabetes using dental variables alone or in combination with other predictive variables, the objectives of the studies varied. Studies by Holm et al. (2016) and Lalla et al. (2011Lalla et al. ( , 2013  (n = 4) risk of bias and probable (n = 1) to uncertain (n = 6) applicability across studies, since 6/7 studies presented only outcomes of modeling without publishing the final models and only 1/7 models was further validated.
Notably, periodontitis-associated measures modeled as risk factors were consistently retained in the predictive models, thus further reinforcing the evidence base supporting bidirectional interaction between undiagnosed T2DM/prediabetes and the infectious/ inflammatory processes associated with periodontitis. Two variables: "PPD" and/or "number of missing teeth" were retained in all seven models as predictors in all populations in which they were tested.
Apart from trauma, or tooth loss associated with extractions associated with either oral oncology-related treatment or caries, periodontitis represents the most likely cause of missing teeth and is, therefore, a surrogate for prior history of periodontitis. Periodontitis-associated tooth loss is attributable to loss of structural support due to damage of surrounding gums and underlying bone caused by chronic periodontal infection and host immune response to the local infectious processes in the gingival tissue. Consistent retention of these two related periodontal variables suggests high generalizability since they were identified by all seven articles across highly variable study populations and dental care settings.
Further, studies that modeled oral variables independently as the main predictors of risk for undiagnosed T2DM/prediabetes found these variables to be highly associated with biological outcomes. Analyses of combined data sets from two studies by Lalla et al. (2011Lalla et al. ( , 2013, reported high sensitivity (0.87) and area under the receiver operating curve (AUC) (0.83) with use of an algorithm including only two optimally defined oral variables: ≥26% of teeth with deep pockets (defined as ≥5 mm and ≥four missing teeth) and glycemic measure (HbA1C) collected at the POC.
As shown in Table 1, the performance of models incorporating oral health-related variables achieved AUC measures of up to 0.92.
Model performance was further enhanced following the inclusion of biological glycemic screening measures as predictive variables during modeling. These results suggest that such predictive models merit further investigation and evaluation as potential tools for diabetic risk prediction in the dental setting.  Hegde et al., 2019). While risk prediction modeling is data-driven and noninvasive, studies by Lalla et al. (2011Lalla et al. ( , 2013 further demonstrated enhanced predictive power (increase in AUC) when oral predictive variables were modeled in conjunction with biological measures of glycemic status as compared to models that did not use oral health variables.

| Limitations
Limitations of the current study included the inability to conduct meta-analysis based on high degree of variability across study populations, outcomes, and study designs. Further, the ability to conduct literature searches in other databases and languages may have resulted in lack of capture and inclusion of all articles meeting inclusion criteria for this review.

| CONCLUSION
This systematic review defined oral health variables with attributable risk for undiagnosed T2DM/prediabetes. Notably, some studies F I G U R E 3 Oral health-related variables identified by the systematic review. This figure summarizes the 19 oral health variables that were evaluated among articles systematically reviewed. Two variables consistently identified and retained in models by all seven articles systematically reviewed included the number of missing teeth and presence of periodontitis based on documentation of variables associated with pathophysiological manifestations of periodontitis demonstrated that oral health variables contributed to models predicting risk undiagnosed T2DM and prediabetes both independently and in combination with biological glycemic measures. The most consistently informative variables across studies included number of missing teeth and demonstration of periodontitis, defined by PPD and CAL. Findings of the current systematic review reflect those of a recently published systematic review of field trials that undertook screening for undiagnosed T2DM/prediabetes at POC in dental settings (Glurich, Bartkowiak, et al., 2018) that collectively reported high prevalence of undiagnosed T2DM/prediabetes across various populations applying POC testing alone or in combination with predictive modeling.