Abstract
A change in the American Diabetes Association guidelines added hemoglobin A1c (HbA1c) to the assays for diabetes diagnosis, but evidence suggests that glucose vs. HbA1c criteria may identify different segments of the affected population. We previously demonstrated that oral findings offer an opportunity for the detection of undiagnosed abnormal fasting plasma glucose (FPG) among dental patients who present with diabetes risk factors. In this new cross-sectional study, we sought to extend these observations. The first goal, using data from 591 new participants, was to assess our previously identified hyperglycemia detection models when HbA1c is used for case definition. The second goal, using data from our total cohort of 1,097 participants, was to evaluate the models’ performance regardless of whether an FPG or an HbA1c is used for diagnosis. The presence of ≥ 26% teeth with deep pockets or ≥ 4 missing teeth correctly identified 72% of pre-diabetes or diabetes cases in the HbA1c sample and 75% in the total population. The addition of a point-of-care HbA1c ≥ 5.7% increased correct identification to 87% and 90%, respectively. These results demonstrate the validity of our prediction models regardless of the test used for diabetes or pre-diabetes diagnosis in the clinical setting and underscore the contribution dentists can make.
Keywords
Introduction
Type 2 diabetes, a serious health condition and an established risk factor for periodontitis, is usually silent in its early stages and often remains unrecognized for years. Recent estimates suggest that one out of four affected individuals in the United States (US) is undiagnosed (Centers for Disease Control and Prevention, 2011). Data from randomized trials indicate that early diagnosis and aggressive treatment are of the essence in secondary prevention of the disease’s many life-threatening complications (American Diabetes Association, 2013a,b).
Pre-diabetes often precedes type 2 diabetes, is characterized by mild hyperglycemia and insulin resistance, and confers a significant risk for vascular disease (Diabetes Prevention Program Research Group, 2007). Progression to diabetes is not inevitable and can be reduced through diet and exercise. Early identification of persons with pre-diabetes, now estimated to number 79 million in the United States (Centers for Disease Control and Prevention, 2011), allows for intervention to reduce the risk (Yamaoka and Tango, 2005; Jeon et al., 2007).
Several diabetes risk assessment models using indices available to primary care physicians have been proposed in the medical literature, and it has been suggested that a two-stage process could be clinically efficient: Non-invasive variables (with little effort and cost) are used first, and laboratory blood tests that have been shown to improve results of non-invasive models can be applied in patients with an increased risk at first assessment for more precise risk estimation (Tabák et al., 2012).
We previously developed and assessed predictive models based on oral findings that can help identify unrecognized abnormal fasting plasma glucose (FPG) levels indicative of pre-diabetes or diabetes among at-risk dental patients (based on their age and self-reported risk factors) (Lalla et al., 2011). Since then, the American Diabetes Association (ADA) introduced the most significant change to diabetes diagnosis in decades by proposing a hemoglobin A1c (HbA1c) ≥ 6.5%, following venipuncture and laboratory analysis by high-performance liquid chromatography (HPLC) but not a point-of-care test, as an alternative diagnostic criterion for diabetes (American Diabetes Association, 2013b). According to the guidelines, a positive test must be confirmed on a second occasion. For pre-diabetes diagnosis, the ADA endorses an HbA1c in the range of 5.7% to 6.4% and no confirmatory testing. FPG and HbA1c levels reflect different aspects of glucose metabolism, and accumulating evidence suggests that glucose- vs. HbA1c-based criteria may identify different segments of the population that overlap only partially, posing a potential challenge for practicing physicians (Boronat et al., 2010; Christensen et al., 2010; Jørgensen et al., 2010; Ackermann et al., 2011; Balkau et al., 2011; Silbernagel et al., 2011; Rathmann et al., 2012).
With this new clinical reality in mind, our first goal in the present study with 591 new participants was to assess whether oral findings have a role in diabetes and pre-diabetes identification among dental patients who present with diabetes risk factors when HbA1c is used for case definition. Our second goal, using our total cohort of 1,097 participants, was to assess the performance of our predictive models regardless of whether an FPG or an HbA1c is eventually used for diagnosis.
Materials & Methods
Participant Recruitment and Examination Protocol
The study protocol, previously described in detail (Lalla et al., 2011), was approved by the Columbia University Institutional Review Board and conforms to STROBE guidelines. Briefly, from April 2010 to June 2012, we obtained written informed consent from 663 individuals among new patients at the College of Dental Medicine who had never been told that they have diabetes or pre-diabetes. To target those at some level of risk for hyperglycemia, we further selected those who: (a) were ≥ 40 yrs old, if non-Hispanic white and ≥ 30 yrs old, if Hispanic or non-white (Dallo and Weller, 2003); and (b) had at least one additional self-reported diabetes risk factor, such as diabetes in first-degree blood relatives, hypertension, hypercholesterolemia, or overweight/obesity.
A single calibrated examiner performed a periodontal evaluation on the 591 participants who met all entry criteria and recorded numbers of missing teeth, probing depths, and presence of bleeding on probing at 6 sites per tooth, excluding third molars. Participants then received a fingerstick point-of-care HbA1c by means of a benchtop analyzer (DCA 2000 Plus; Bayer HealthCare, Tarrytown, NY, USA). This test provides rapid results which can be used for patient feedback at the time of encounter, a practice that has been shown to improve long-term glycemic control in diabetic patients (Cagliero et al., 1999). A venipuncture sample was collected at the end of the visit for a diagnostic HPLC HbA1c test and was analyzed at the New York Presbyterian Hospital/Columbia University Medical Center Special Chemistry Laboratory.
Case Definitions and Analysis
Data from the new cohort were analyzed and are presented, both separately and together with the 506 participants from our previous report (Lalla et al., 2011). The latter were recruited from April 2009 to March 2010, at the same clinic, according to the same criteria, but returned following an overnight fast and were assessed by an FPG diagnostic test. The same examiner performed the screening/examination of all participants in both cohorts. Hyperglycemia was defined according to the ADA criteria as follows: diabetes, if FPG ≥ 126 mg/dl or HPLC HbA1c ≥ 6.5%; pre-diabetes, if FPG = 100-125 mg/dl or HPLC HbA1c = 5.7-6.4%. All patients were informed of their results, and those with an abnormal outcome were advised to see a physician for further testing and appropriate care.
Statistical Analyses
The missing data rate was very low (0-2%), and the “completer” or “complete case” analysis approach to managing missing data (deleting all participants with incomplete data in the variables involved from the analysis) was used. Continuous variables were summarized and are presented as mean ± SD, with categorical variables as count (percentage). Analysis of variance (ANOVA) tests were used to compare continuous variables, and Chi-square tests were used to compare categorical variables among the three groups (normal HbA1c, pre-diabetes, and potential diabetes). Post hoc t tests were used for pair-wise comparisons. Conventional receiver operating characteristic (ROC) curve analysis was conducted using multiple logistic regression models, and ROC plots were generated with R software (version 2.8.1, R Foundation for Statistical Computing, Vienna, Austria). The areas under the curves (AUCs) were compared by the DeLong, DeLong, and Clarke-Pearson test for measurement of the models’ ability to discriminate between normal and abnormal HbA1c/FPG (binary classification). For measurement of the models’ ability to differentiate among the 3 conditions (i.e., health vs. pre-diabetes vs. diabetes) simultaneously, a three-class ROC analysis was performed (Mossman, 1999). The U-statistic was used to estimate the volume under the surface (VUS), and 100 bootstrap samples were used to construct the 95% confidence intervals by the method of Li and Fine (a VUS of 0.167 corresponds to random prediction) (Li and Fine, 2008). All other analyses were conducted with SAS, version 9.2 (SAS Institute, Cary, NC, USA). All p values ≤ .05 were considered statistically significant.
Results
Among the 591 newly recruited participants who received the HbA1c diagnostic testing, 326 were identified with an abnormal result: 51 (8.6%) as potentially diabetic, and 275 (46.5%) as pre-diabetic. Information on age, gender, race, ethnicity, periodontal parameters, presence of risk factors, smoking status, and point-of-care HbA1c levels for the normal HbA1c and the two abnormal HbA1c groups (pre-diabetes and diabetes) is summarized in Table 1. Individuals in the normal HbA1c group had a mean age of 50.9 ± 11.8 yrs vs. 56.7±11.6 yrs in the pre-diabetes group and 58.8 ± 11.7 in the potential diabetes group. Numbers of missing teeth and percentage of teeth with at least one deep (≥ 5 mm) pocket were significantly higher in the diabetes group, compared with both the healthy and pre-diabetes groups.
Participant Characteristics by HbA1c Test Outcome (N = 591)
Normal HbA1c: HPLC HbA1c < 5.7%. Pre-diabetes: HPLC HbA1c 5.7-6.4%. Potential diabetes: HPLC HbA1c ≥ 6.5%. The number of participants with missing data for a given variable is shown in parentheses. Data are shown as mean ± SD or n (%). Overall p values are from an ANOVA F test or Chi-square test, comparing all 3 groups.
Statistically significantly different compared with normal HbA1c; £ statistically significantly different compared with pre-diabetes. aThe vast majority of our Hispanic participants refused to self-identify with any of the racial groups, choosing the option “other” or “unknown”. bOut of 28 teeth.
The 2 predictive models we had previously singled out for their discriminative power in detecting abnormal FPG (Lalla et al., 2011) were assessed in this cohort, by calculating the AUC using multiple logistic regression. The model including only percentage of teeth with at least one pocket ≥ 5 mm and number of missing teeth (model A) had an AUC of 0.58 (95% CI: 0.531, 0.624). The addition of a point-of-care HbA1c result to model A significantly improved the AUC (model B, AUC: 0.92, 95% CI: 0.904, 0.944; p < .0001). As expected, the point-of-care HbA1c test alone had an AUC of 0.92 (95% CI: 0.903, 0.943), significantly better than that of model A (p < .0001), and no different from that of model B (p = .47).
Using the previously identified optimal cut-offs (26% for the percentage of teeth with deep pockets, 4 for the number of missing teeth, and 5.7% for the point-of-care HbA1c) for the variables included in models A and B (Lalla et al., 2011), we calculated the models’ performance for predicting abnormal HPLC HbA1c (Table 2). The performance characteristics shown in Table 2 indicate that the presence of at least 26% teeth with deep pockets or at least 4 missing teeth can correctly identify 72% of true cases with previously unrecognized hyperglycemia in this cohort; as expected, the addition of a point-of-care HbA1c result ≥ 5.7% increases correct identification to 87%.
Sensitivity, Specificity, Positive Predictive Value (PPV), and Negative Predictive Value (NPV) in the Detection of Pre-diabetes and Diabetes (Defined by HPLC HbA1c) Using Previously Identified (Lalla et al., 2011) Optimal Cut-off Values (N = 591)
Further, and since either an FPG or an HPLC HbA1c may be used clinically by a physician for diagnosis of pre-diabetes/diabetes, we performed analyses for the total population: 591 with the HbA1c outcome from the current study and 506 with the FPG outcome from our previous report (Lalla et al., 2011). The two cohorts differed in the diagnostic test used to define dysglycemia, but were recruited in the same clinic, with the same criteria, assessed by the same examiner, and displayed similar overall characteristics. Out of the total population of 1,097 individuals, 436 (39.7%) were identified as pre-diabetic and 72 (6.6%) were identified as potentially diabetic. Model A (% of teeth with at least one deep pocket and number of missing teeth) had an AUC of 0.60 (95% CI: 0.562, 0.630). Model B (model A + point-of-care HbA1c) had a significantly better AUC of 0.83 (95% CI: 0.800, 0.849; p < .0001). ROC curves for these 2 models are depicted in the Fig. The AUC for the point-of-care HbA1c test alone was 0.82 (95% CI: 0.798-0.846). Performance characteristics for the whole population, again with the cut-offs identified in our previous report, are shown in Table 3.
Receiver operating characteristic (ROC) curves for the previously identified (Lalla et al., 2011) prediction models using logistic regression. All participants, N = 1,097. Model A: % of teeth with at least one deep (≥ 5 mm) pocket + number of missing teeth. Model B: % of teeth with at least one deep (≥ 5 mm) pocket + number of missing teeth + point-of-care (POC) HbA1c. Model A had an area under the curve of 0.60 (95% CI: 0.562, 0.630). Model B had a significantly better area under the curve of 0.83 (95% CI: 0.800, 0.849; p < .0001).
Sensitivity, Specificity, Positive Predictive Value (PPV), and Negative Predictive Value (NPV) in the Detection of Pre-diabetes and Diabetes (Defined by Either FPG or HPLC HbA1c) Using Previously Identified (Lalla et al., 2011) Optimal Cut-off Values (N = 1,097)
Further, the large number of participants in the whole cohort allowed us, for the first time, to assess separately the performance of the 2 models in differentiating individuals with frank diabetes. Model A had an AUC of 0.64 (95% CI: 0.579, 0.711) for diabetes identification and, with the predefined cut-offs, sensitivity of 0.82 and specificity of 0.32. Model B had a significantly better AUC of 0.96 (95% CI: 0.947, 0.980), a sensitivity of 1.00, and a specificity of 0.21. As a reference, the point-of-care HbA1c test alone had an AUC of 0.96 (95% CI: 0.946, 0.979), a sensitivity of 1.00, and a specificity of 0.59.
Finally, to assess the ability of the models to detect and characterize hyperglycemia (i.e., to discriminate each of the 3 potential outcomes: health vs. pre-diabetes vs. potential diabetes), we performed a three-class ROC analysis using the whole population. We found that model A had a VUS of 0.21 (95% CI: 0.160-0.250), which suggests an improved performance compared with a random estimate (VUS of 0.17) and a difference approaching statistical significance. Model B had a VUS of 0.57 (95% CI: 0.482-0.685), indicating a stronger performance in the three-group classification task and a statistically significant improvement compared with the random estimate. The VUS for the point-of-care HbA1c test alone was 0.58 (95% CI: 0.503-0.659).
Discussion
For decades, levels of blood glucose following either fasting or an oral glucose load have formed the sole basis for diabetes diagnosis, but a recent change in the ADA diagnostic criteria introduced HPLC HbA1c as a diagnostic tool for the first time (American Diabetes Association, 2013b). The findings of the current study in a large population of dental patients demonstrate that our undiagnosed hyperglycemia prediction models perform well, regardless of whether an FPG or an HbA1c test is used for case definition. Thus, and since a physician in a clinical setting may use either test for pre-diabetes and diabetes diagnosis, these findings lend further credibility to the use of our simple screening approaches by dental professionals.
The two-dental-variable model has a sufficient ability to discriminate between 2 groups i.e., health vs. pre-diabetes and diabetes (AUC = 0.60), or health and pre-diabetes vs. diabetes (AUC = 0.65). The need to identify individuals with frank diabetes is crucial, but the benefits from early identification and treatment of those with pre-diabetes are also becoming increasingly apparent (Gong et al., 2011; Perreault et al., 2012). Evidence has suggested that pre-diabetes identification warrants patient education regarding modifiable risk factors and lifestyle changes to mitigate risk (Inzucchi, 2012; American Diabetes Association, 2013a). It is also proposed that pre-diabetes should be treated not only to prevent progression to diabetes, but also to prevent the potential effects of pre-diabetes itself (Tabák et al., 2012). When faced with the more complicated three-group classification task (i.e., health vs. pre-diabetes vs. diabetes), the two-dental-variable model does not perform as strongly as in the binary classification charge. A point-of-care HbA1c test result clearly has value in the three-group classification task, but other variables/risk factors may need to be combined with the periodontal parameters to improve the model’s ability not only to detect, but also to characterize the level of hyperglycemia; this is an important area for future investigation.
A two-stage identification approach, similar to that suggested for primary care physicians (Tabák et al., 2012), could also be proposed for dental professionals. First, dental variables are considered in those patients who present with other easily identifiable risk factors. Then, in those patients with an increased risk at first assessment, a point-of-care blood test can be applied, if available to the dentist and acceptable by the patient. Importantly, our suggested simple models have high sensitivity, similar to that reported for several diabetes risk assessment approaches studied in medical settings (Lin et al., 2009). We believe that sensitivity is an important performance characteristic for screening in the dental setting and that false-positive results are not critical, since the dental professional will not deliver a diagnosis or treat the patient’s hyperglycemia.
Interestingly, no diabetes prediction model has been universally accepted for use in medical settings, and it has been suggested that, given that race/ethnicity is strongly related to diabetes risk, recalibration of algorithms might be necessary when models are applied to different populations (Tabák et al., 2012). A limitation of the current effort is that our algorithms were developed and assessed in a Northern Manhattan population that is predominantly Hispanic. Testing to assess the external validity of these algorithms in diverse patient populations is an important future effort.
Since our 2011 report on the simple models that could be used for the identification of pre-diabetes and diabetes among dental patients (Lalla et al., 2011), work from other investigators has supported the notion of screening for undiagnosed diabetes in dental settings and, in addition, provided evidence that implementation in dental practices is feasible and that most patients and dental providers believe that the dental visit is a good opportunity for early diabetes identification (Barasch et al., 2012; Rosedale and Strauss, 2012). Taken together, the current results further underscore that dental professionals have the opportunity to assume an active role in identifying, among their patients who present with diabetes risk factors, those with undiagnosed hyperglycemia and direct them to receive appropriate medical evaluation and care. Given the legacy effect of glycemic control on the development of oral and systemic complications, both early glycemic stability in those affected by diabetes who remain unrecognized and a delay (even a modest one) in the onset of diabetes among those with pre-diabetes are valuable goals.
Footnotes
Acknowledgements
The authors thank Awilda Marte, Richard Buchsbaum, and Andrew Ferraro (all at Columbia University) for their valuable assistance.
The study was supported by research grants from Colgate Palmolive and the New York State Health Foundation. The funders had no role in the study design, data collection and analysis, or preparation of the manuscript.
The authors declare no potential conflicts of interest with respect to the authorship and/or publication of this article.