Abstract
Aims
There are numerous risk or screening scores for the prediction of type-2 diabetes mellitus (DM). In contrast, few scores are available for preDM. In this paper, we compare the two screening scores from the American Diabetes Association (ADA) and Centers for Disease Control and Prevention (CDC) that can be used for DM as well as preDM.
Methods
Adult participants (N=9,391) without known DM from the National Health and Nutrition Examination Surveys 2009–12 were included. We fitted the factors/items in the ADA and CDC scores in logistic regression with the outcomes of undiagnosed DM, preDM, and combination, and assessed the association and discrimination accuracy. We also evaluated the suggested cutpoints that define high risk individuals. We mimicked the original models/settings but also tested various deviations/modifications often encountered in practice.
Results
Both scores performed well and robustly, while the ADA score performed somewhat better (e.g., AUC=0.77 for ADA and 0.73–0.74 for CDC for DM; 0.72–0.74 and 0.70–0.71 for preDM). The same predictors and scoring rules seem to be reasonably justified with different cutpoints for DM and preDM, which can make usage easier and consistent. Some factors such as race and HDL/LDL cholesterols may be useful additions to health education.
Conclusions
Current DM education and screening focus on the prevention and management of DM. The ADA and CDC scores could further help when we identify individuals at high risk for preDM, and teach the importance of preDM during which lifestyle intervention can be effective and urgently needed.
Keywords: ADA, CDC, prediabetes, risk assessment, risk score, screening
There are a number of prediction or screening scores/models for incident and prevalent type-2-diabetes-mellitus (DM) worldwide (http://www.idf.org/epidemiology/risk-prediction). Some are actively being utilized in clinical and community settings or for research purposes, say, for self-assessment, health education and patient-doctor communication/shared decision making. In contrast, there are few screening scores for preDM, and some may question whether we need scores for preDM, different from those for DM. To our knowledge, the two scores from the Centers for Disease Control and Prevention (CDC) and the American Diabetes Association (ADA) that have been developed to help screening DM as well as preDM are relatively well known and easy to use (say, in the pencil-and-paper questionnaire): namely, ‘CDC prediabetes screening test’, http://www.cdc.gov/diabetes/prevention/pdf/prediabetestest.pdf and the ‘ADA diabetes risk test’, http://www.diabetes.org/are-you-at-risk/diabetes-risk-test/. The original models for these scores were developed for the outcome of undiagnosed DM from the National Health and Nutrition Examination Survey (NHANES) 2004 or earlier, by statistical modeling [1 2].
Specifically, the ADA score consists of 7 questions (total score of 0–11) on age, sex, gestational DM, family history of DM, hypertension, physical activity, and obesity (based on body mass index (BMI) via a weight-height chart). The CDC score consists of 7 questions on 6 factors (total score of 0–18) based on age, having delivered a baby weighing more than 9 pounds, sibling’s DM, parent’s DM, physical activity, and obesity; see the scoring algorithms in the Figure S1. Although the original scores were developed to identify individuals at elevated risk for undiagnosed DM, they were also suggested to be used for undiagnosed preDM, with different cutpoints: ≥ 5 for DM and 4 for preDM in the ADA score and ≥ 10 for DM and 9 for preDM in the CDC score [1–3]. We also found that some modifications or adaptations are often accompanied to handle realistic issues or improve uptake (e.g., related to data unavailable or limited, less user-friendly questions, varying definitions).
In this paper, we evaluated these two scores in terms of prediction/detection of the outcomes – DM; preDM; and DM and preDM combined, all undiagnosed – and if we can support the use of the same score with different cutpoints for DM and preDM. We also conducted sensitivity and exploratory analyses in order to assess the robustness of the models’ performance under various modifications/deviations (e.g., in defining or understanding variables) and restrictions (e.g., on age groups), and the value of additional risk factors commonly considered in relevant contexts. This study may provide some lessons to practitioners, researchers, educators, and users regarding how to wisely use good diabetes and other risk assessment tools in practice.
METHODS
Survey Design and Participants
We used the NHANES 2009–12, the most recent waves at the time of the study. We restricted our analyses to the adult population, who are ≥ 20 years old. We excluded individuals with 1) diagnosed DM (i.e., doctor told you or currently on DM medication) or 2) missing outcomes data (i.e., fasting glucose, A1C, and 2-hour plasma glucose by oral glucose tolerance test (OGTT) unmeasured). In the analyses where preDM is the sole outcome, we further excluded those with undiagnosed DM and diagnosed preDM (e.g., doctor told you). Publicly available data were used in our study (http://www.cdc.gov/nchs/nhanes.htm).
Outcomes and predictors
We focused on the variables that are needed in the derivation or use of the two screening scores. We defined predictors and outcomes following the original definitions or the current practice guidelines [4 5] as closely as possible in the primary analyses. Some modifications/adaptations were addressed in the sensitivity/ancillary analyses. To reflect the most common scenario, if data on risk factor is missing, we assigned the score of 0, so we equated the answers of ‘No’ and ‘I don’t know’.
The outcomes of type-2 DM and preDM are defined as follows: If a person has fasting glucose ≥ 7.0 mmol/L, A1C ≥ 48 mmol/mol, or 2-hour glucose ≥ 11.1 mmol/L, then this person has DM. If a person does not meet the DM criteria, but has 5.6 ≤ fasting glucose<7.0, 39 ≤ A1C<48, or 7.8 ≤ 2-hour glucose<11.1, then this person has preDM.
Predictors are defined in the following manner. Age is categorized with the cutpoints of 40, 50 and 60 for the ADA score and of 45 and 65 for the CDC score. Hypertension is defined based on diagnosis (i.e., told by doctor), medication use, or blood pressure (systolic ≥ 140 mmHg or diastolic ≥ 90 mmHg using the higher value of the first two measurements). Family history of DM is defined based on parent and sibling’s DM. [Of note, NHANES we used did not collect family history information separately for parent and sibling so we combined the 2 questions into 1 in the CDC score and assigned the score of 1 in the main analyses. We also assigned the score of 2 and a combination of 1 and 2 in sensitivity analyses.] Pregnancy history data were available so we coded as Yes/No. We created obesity categories as specified in the two scores. The paper version of the both scores provides a small table of weight and height, where the classification corresponds to BMI cutpoints of 25/30/40 for the ADA score (4 groups) and of 27 for the CDC score (2 groups). Finally, there are numerous ways to assess physical activity. The CDC score asks “Get little or no exercise in a typical day?” and the ADA score askes “Are you physically active?” but the same questions were not utilized in the NHANES. Considering these and currently available recommendations from the ADA and CDC (http://www.diabetes.org/are-you-at-risk/lower-your-risk/activity.html & http://www.cdc.gov/diabetes/living/beactive.html), we derived a binary variable by checking if 5 or more days in a typical week of any of the following activities: vigorous or moderate work, recreational work, walk or bicycle.
We described the variables used in sensitivity and ancillary analyses in Appendix. We tried to address frequently encountered situations in a variety of realistic settings where risk scores are used.
Statistical Analyses
We combined the NHANES 2009–2010 and 2011–2012 and accounted for complex survey design in relevant analyses according to the NHANES’s analytic guidelines. We repeated some analyses with different weights (e.g., medical exam weight in place of fasting subsample weight) or no weight to include maximum sample/information available, where these 3 different weighting schemes may achieve lowest bias and higher efficiency in estimation, and closeness to real practice (like a convenient sample in community screening). We indicated which weight was used in each analysis in tables’ footnotes. We described the study participants by summary statistics and computed the prevalence of undiagnosed DM and preDM for each total score of the ADA and CDC scores.
We fitted logistic regression with the predictors and each outcome described above, and computed odds ratio (OR) with 95% confidence interval (CI) and p-value for quantifying associations, and the area under the receiver-operating-characteristic curve (AUC) for assessing discrimination ability. In addition, we calculated standard performance measures – the percent of high risk individuals, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV).
We performed sensitivity and subgroup analyses of practical importance, and tested the predictiveness of additional factors that are not included in the two scores. In these exploratory analyses, we analyzed ordinal and continuous variables as continuous predictors and nominal and binary variables as categorical predictors in regression. Analyses were conducted in SAS 9.3 (SAS Institute, Cary, NC).
RESULTS
The characteristics of 9,391 participants included in our study, excluding diagnosed DM, are described in Table 1. The average age was 46 years [range: 20 to >80] and 48% were men. Obesity status based on BMI showed a higher proportion of ‘higher than normal weight’, compared to that based on self-report (67 vs. 57%). Of note, the NHANES did not allow ‘obese’ as an answer to the question “How do you consider your weight?” A total of 45% of participants reported they were physically active for 5 or more days of a week. Approximately 7% were shown to have undiagnosed DM and 48% to have undiagnosed preDM under the optimal setting (e.g., 3 DM diagnostic tests were performed, and the best-suited weights were used for estimation). The prevalence varies depending on the tests available and analysis samples/weights/methods used [6]. Nonetheless, only 4.5% of participants reported they were told by doctor they had preDM.
Table 1.
Predictors included in the ADA and CDC scores | Mean (standard error) or Percentage |
---|---|
Age, year | 45.9 (0.48) |
Men | 47.8% |
Body mass index, kg/m2 | 28.3 (0.12) |
25–30 (overweight) | 34.3% |
30–40 (obese) | 27.3% |
≥ 40 (extreme obese) | 5.2% |
Overweight (self-report) | 57.3% |
Waist circumference, cm - Women | 94.7 (0.36) |
Men | 100.0 (0.45) |
Systolic blood pressure, mmHg | 122.9 (0.38) |
Diastolic blood pressure, mmHg | 73.0 (0.39) |
Hypertension | 34.8% |
Family history of DM | 32.8% |
Physically active | 45.0% |
Having a baby >9 pounds (women only) | 11.4% |
Gestational diabetes (women only) | 3.9% |
Outcomes-related | |
Fasting glucose, mmol/L, n=4290 | 5.5 (0.02) |
A1C, % or mmol/mol, n=9380 | 5.5 (0.01) or 36.3 (0.10) |
2-hour glucose, mmol/L, n=3883 | 6.4 (0.05) |
Diabetes status: Normal/preDM/DM by fasting glucose only |
57.5/38.9/3.6% |
by A1C only | 72.7/25.0/2.3% |
by 2-hour glucose only | 80.8/13.8/5.5% |
by glucose/A1C/2-hour glucose* | 45.4/47.7/6.9% |
PreDM told by doctor | 4.5% |
Adult participants are those of 20 years old or older.
Those with missing outcomes data (i.e., all of fasting glucose, A1C and 2-hour glucose) were excluded.
Fasting glucose and 2-hour glucose were collected from subsamples.
All summary statistics were weighted with interview, medical exam, fasting or oral glucose tolerance test (OGTT) sample weight.
Sample sizes, total N and n, were unweighted.
OGTT weight was used to estimate the prevalence more validly, following the NHANES analytic guidelines. When medical exam weight or no weight was used, which can include larger N, prevalence turns out to be 60.0/35.5/4.5% and 55.0/38.5/6.4%, respectively.
DM denotes type 2 diabetes mellitus.
When we fitted the logistic regression for the outcome of undiagnosed DM, which was the primary outcome in the original studies [1 2], all predictors except for ‘macrosomic baby (>9 pounds)’ were statistically significant, with the AUC of 0.77 for the ADA score and 0.73–0.74 for the CDC score. When we fitted the same set of predictors for the outcome of undiagnosed preDM (N=8,442 after excluding undiagnosed DM and diagnosed preDM), the observed ORs were attenuated toward the null, with the directions of the association being preserved. The AUC decreased to 0.72–0.74 and 0.70–0.71, respectively, for the ADA and CDC score, which is anticipated as preDM includes a wider range of patients away from the tail of the risk spectrum; see Table 2. When we combined DM and preDM as composite outcome, the corresponding AUCs slightly increased (0.73–0.76 and 0.71–0.72), also as anticipated. Figure S2 demonstrates the increasing trend of the disease prevalence as the total score increases.
Table 2.
a. Outcome=Undiagnosed DM (N=9391) | |||||
---|---|---|---|---|---|
ADA (AUC=.77/.77/.77) | CDC (AUC=.73/.74/.73) | ||||
Predictor | OR (95% CI) | p-value | Predictor | OR (95% CI) | p-value |
Age 40–49 | 2.3 (1.5–3.6) | <0.001 | Age<65 & physically inactive | 1.6 (1.2–2.1) | <0.001 |
50–59 | 3.5 (2.3–5.4) | <0.001 | 45–64 | 2.8 (2.2–3.6) | <0.001 |
≥60 | 6.3 (4.3–9.3) | <0.001 | ≥65 | 8.0 (5.6–11.6) | <0.001 |
Men | 1.5 (1.1–2.0) | 0.007 | |||
Hypertension | 1.5 (1.3–1.9) | <0.001 | |||
Family DM | 1.7 (1.3–2.1) | <0.001 | Parent DM Sibling DM |
1.8 (1.4–2.3) | <0.001 |
BMI 25–29.9 | 1.5 (1.0–2.2) | 0.035 | BMI≥27 | 2.8 (2.1–3.6) | <0.001 |
30–39.9 | 3.1 (2.1–4.7) | <0.001 | |||
≥40 | 7.3 (5.0–10.7) | <0.001 | |||
Physically inactive | 1.5 (1.2–1.9) | <0.001 | Physically inactive | See “age” above | |
Gestational DM (women only) | 2.5 (1.3–5.0) | 0.006 | >9 pound baby (women only) | 1.0 (0.6–1.5) | 0.92 |
b. Outcome=Undiagnosed preDM (N=8442) | |||||
---|---|---|---|---|---|
ADA (AUC=.72/.74/.72) | CDC (AUC=.70/.71/.70) | ||||
Predictor | OR (95% CI) | p-value | Predictor | OR (95% CI) | p-value |
Age 40–49 | 2.1 (1.7–2.5) | <0.001 | Age<65 & physically inactive | 1.0 (0.9–1.1) | 0.96 |
50–59 | 2.8 (2.3–3.4) | <0.001 | 45–64 | 2.7 (2.3–3.1) | <0.001 |
≥60 | 5.0 (4.1–6.2) | <0.001 | ≥65 | 4.9 (4.1–5.9) | <0.001 |
Men | 1.5 (1.3–1.7) | <0.001 | |||
Hypertension | 1.3 (1.1–1.4) | <0.001 | |||
Family DM | 1.2 (1.0–1.3) | 0.01 | Parent DM Sibling DM |
1.2 (1.1–1.4) | 0.002 |
BMI 25–29.9 | 1.4 (1.2–1.7) | <0.001 | BMI≥27 | 1.9 (1.7–2.3) | <0.001 |
30–39.9 | 2.1 (1.7–2.6) | <0.001 | |||
≥40 | 3.4 (2.6–4.5) | <0.001 | |||
Physically inactive | 1.0 (0.9–1.1) | 0.82 | Physically inactive | See “age” above | |
Gestational DM (women only) | 2.0 (1.3–3.0) | 0.002 | >9 pound baby (women only) | 1.0 (0.8–1.3) | 0.91 |
Regression fits were from weighted analyses by medical exam weight. Three AUCs are from regressions with medical exam weight/fasting weight/unweighted, respectively. BMI=body mass index.
In b, undiagnosed DM and diagnosed preDM were excluded. When we combined the outcomes (undiagnosed preDM and undiagnosed DM), AUC=.73/.76/.73 for the ADA score and .71/.72/.71 for the CDC score. DM denotes type 2 diabetes mellitus.
When different cutpoints were evaluated for different outcomes, sensitivity was highest when these scores were used for the identification of DM, 0.83 for the ADA score and 0.79 for the CDC score. When we aimed at the identification of preDM (after excluding undiagnosed DM artificially), sensitivity was somewhat lowered, but PPV markedly increased (e.g., ~0.10 to >0.50), which is not unexpected as PPV depends on the disease prevalence. The performance of these scores was slightly enhanced when it was used to identify DM and preDM together – rather than preDM alone – which could reflect a more realistic scenario in practice because persons with undiagnosed DM or preDM are unaware of their condition so eligible to use the score; see Table 3. As noted before, the CDC score includes parent’s DM and sibling’s DM as separate predictors but the NHANES did not collect these variables separately. Thus, we tried 4 scenarios: 1) assign the score of 1 for the family history of DM; 2) assign the score of 2; 3) assign the score of 2 or 1, where 2 to 25% of those who had DM; and 4) repeat the third scenario but replacing 25% by 50%. These 4 experiments yielded the identical results.
Table 3.
Model/Cutpoint | Outcome, undiagnosed | % of high risk | Sensitivity | Specificity | PPV | NPV |
---|---|---|---|---|---|---|
ADA ≥ 4 | preDM/DM | 60 | 78 | 54 | 57 | 76 |
≥4 | preDM | 58 | 76 | 54 | 53 | 77 |
≥5 | DM | 46 | 83 | 57 | 12 | 98 |
CDC ≥ 9 | preDM/DM | 58 | 74 | 54 | 56 | 73 |
≥9 | preDM | 56 | 72 | 54 | 51 | 74 |
≥10 | DM | 52 | 79 | 50 | 10 | 97 |
Analyses were unweighted, assuming we have convenient sample in a community screening setting.
The CDC model yielded the same results with score of 1, 2, or a combination of 1 or 2 for family DM.
For outcome=preDM, undiagnosed DM and diagnosed preDM cases were excluded.
For outcome=preDM/DM, diagnosed preDM cases were excluded.
In all analyses, diagnosed DM cases were excluded.
DM denotes type 2 diabetes mellitus; PPV=positive predictive value; NPV=negative predictive value.
When we introduced various modifications on variables’ definitions, the AUC values were quite robust, which may justify some modification(s) are acceptable (Table 4). We observed that fasting glucose alone in the outcome definition yielded the lowest AUCs. Discrimination ability of the scores was consistently higher in younger age group, less than 60 years old. However, when the outcome was preDM, AUCs were the highest when younger and older groups were combined. We found that knowing accurate obesity status in more than two categories seems to be important because when we used a binary status (overweight vs. normal/underweight) based on self-report, the lowest AUC was resulted and regression analyses clearly demonstrated strong monotonic associations in obesity grade and DM as well as preDM risk. Also, when waist circumference (WC) was added to the existing models where BMI-based categories were already in (as in Table 2), WC was highly significant (p’s<0.001). Notably, when BMI vs. WC vs. waist-to-height ratio (WHtR; or waist-to-stature ratio (WSR)) were compared, WHtR yielded the highest AUC, confirming previous findings [7 8]. Among additional predictors tested, race yielded the largest increase in AUC for DM and LDL did for preDM [9 10]. It is inherently difficult to measure the types and amounts of physical activity precisely. Assessment by three different ways led to substantially similar AUC values.
Table 4.
Undiagnosed DM | Undiagnosed preDM | |||
---|---|---|---|---|
ADA | CDC | ADA | CDC | |
Original model | .766 | .731 | .718 | .697 |
Task 1: Different outcome definition | ||||
Fasting glucose alone | .765 | .729 | .717 | .682 |
A1C alone | .785 | .738 | .733 | .720 |
2-hour glucose alone | .768 | .743 | .746 | .729 |
Task 2: Age subgroups | ||||
Age ≥ 60 years old | .637 | .598 | .575 | .577 |
< 60 years old | .774 | .723 | .701 | .666 |
Task 3: Modification in predictor | ||||
Clinical measures: No blood pressures available |
.764 | NA | .717 | NA |
Obesity: Add WC as new predictor |
.773 | .760 | .721 | .709 |
Combine BMI+WC in obesity categories | .767 | NA | .719 | NA |
Over vs. normal/underweight based on self-assessment | .744 | .718 | .705 | .680 |
BMI | .772 | .757 | .719 | .704 |
WC | .774 | .761 | .720 | .709 |
Waist-to-height ratio | .784 | .771 | .724 | .707 |
Physical activity: >30 mins/day of recreational activity |
.767 | .732 | .718 | .698 |
Add ‘hours of sedentary activity’ as new predictor | .766 | .731 | .719 | .700 |
Medical history: No pregnancy data | .762 | .731 | .717 | .697 |
No family DM data | .760 | .722 | .717 | .693 |
Task 4: Additional predictor added | ||||
Race (Black, Hispanic, White, Others) | .779 | .746 | .723 | .705 |
HDL cholesterol | .773 | .746 | .722 | .708 |
LDL cholesterol | .769 | .738 | .738 | .710 |
Total Cholesterol | .766 | .733 | .720 | .702 |
Alcohol (average number of drinks/day) | .776 | .741 | .721 | .697 |
Smoking (current vs. others) | .769 | .738 | .714 | .692 |
(average number of cigarettes/day) | .764 | .752 | .719 | .689 |
Healthy diet (in 5 levels) | .769 | .738 | .719 | .701 |
Total sugar | .768 | .735 | .718 | .700 |
Total fat | .769 | .734 | .718 | .699 |
Predictors were included as continuous variable in logistic regression, unless stated otherwise.
Analyses were weighted with medical exam weight.
Largest AUCs in each model under different tasks are in bold and smallest AUCs are in italicized bold.
Sample sizes of complete data and types of variables (e.g., continuous vs. binary) are different so comparisons need some caution.
DM denotes type 2 diabetes mellitus; NA: not applicable; AUC=under the receiver-operating-characteristic curve; BMI=body mass index; WC=waist circumference.
DISCUSSION
In this study, we evaluated and compared the two preDM screening scores. They were originally derived for undiagnosed DM as an outcome, but have been proposed to use for preDM with different cutpoints. The ADA and CDC scores performed well for DM as well as preDM in independent data, recent NHANES, and we view this as external, temporal validation. The ADA score performed somewhat but nearly uniformly better, and we believe this is partly due to multiple categories used for age and obesity which show strong monotonicity in disease prevalence. Both scores are easy, cheap and noninvasive to administer in the format of pencil-and-paper or online calculator, so either one may be used depending on the preference.
The observed AUC values, the discrimination statistic, for the ADA and CDC scores were comparable to those for well-known risk scores in cardiovascular disease and DM/preDM [11–13], and quite robust when various, small changes were incorporated. For example, when BP or pregnancy-related data were unavailable, which are common in some situations where risk assessment is done (say, with or without interview, or using administrative or health record database), the AUC values were not discernibly reduced. However, our study suggests that it could be important to know the accurate status of a person’s obesity, which supports the inclusion of the existing BMI table in the paper version of these scores, designed to be user-friendly for intended users, say, based on weight (in pound) and height (in inches) in the US. We also observed WC – particularly, WHtR – appears to be more predictive of DM and preDM than BMI. The limitations of BMI are well documented, and some risk scores include WC [8 14–16]. Yet, based on our own and others’ experiences, WC has other issues, for instance, not normally collected in medical record, not easy to measure accurately (often leading to under-estimation), or patient does not know or feel comfortable to be measured [17 18]. Possibly, a currently recommended threshold of 0.5 for WHtR is easy to remember and may carry an educational value. More discussion is warranted regarding how to choose and use anthropometric measures for risk assessment and screening, and for different races or regions/countries [16 19–21]. Until then, weight and height in the screening score, and WC or WHtR in the accompanying educational materials may be ideal.
Our study may have some implications in the development, validation and utilization of risk score. Development of risk score or prediction model is basically dictated by data availability. For example, if the history of gestational DM is not available, researcher cannot include this variable in the model, which is common in the risk scores developed using secondary data. Yet, subjective decision to add gestational DM might be justified with a score assigned in an ad-hoc manner (say, minimal score such as 1), if compelling evidence is available in the literature. A similar issue can happen when data were not systematically measured, which is common when multiple datasets or disparate cohorts are merged [22 23]. On the other hand, it is not always good to include all covariates that are statistically significant and clinically explainable, especially, those with small effect size, costly, less user-friendly or conflicting/controversial variables [24–28]. Moreover, prediction models can be different depending on goal, e.g., patient’s self-assessment vs. shared decision making by patient-doctor vs. policy. Indeed, some arbitrariness/subjectivity in the final predictor selection, score assignment and cutpoint determination was involved in the developments of the ADA and CDC tools [2 29]. The ultimate justification will be tested when the scores will be validated independently with necessary data for the intended goal. As other scientific findings, risk scores can be adapted and evolve naturally whenever sufficient evidence calls for [30]. Furthermore, the same score (with or without small adaptations) may be justified for closely-related conditions; it may help a smaller number of models/scores be developed, so that good models available are used more widely, systematically and wisely, and patients and users become less confused but more comfortable and familiar [16 24 31 32].
Based on our exploratory investigation, diabetes prevention program may emphasize race disparity if that helps the awareness, healthy lifestyle education or more targeted screening; in our study, Hispanics showed the highest (unadjusted) prevalence of undiagnosed DM and Blacks showed the highest prevalence of undiagnosed preDM. The screening scores could be more effective and useful for younger persons, where a high risk group can be recommended to receive blood test or advice from healthcare providers. In addition to well-validated predictors already included in the scores, more emphasis on additional modifiable factors that lay persons can understand easily (e.g., WC, LDL/HDL cholesterols, and diet/nutrition) may be worthwhile.
After excluding known DM, we found that 2–7% of the participants had newly diagnosed DM. 14–48% were shown to have preDM but less than 5% reported doctors told them they had preDM [6]. We think these statistics are alarming. If either screening score were used for preDM in a similar population (e.g., general population in a community in the US), >50% of people would be declared to be at high risk of preDM, and 1 out 2 of them to be revealed to have preDM.
Our study has some limitations. First, the NHANES did not collect parent and sibling’s DM history separately. When we implemented multiple scenarios including conservative and liberal ones, the results were unchanged. Second, sample sizes for variables were different (e.g., fasting and OGTT subsamples, non-response) but some were enforced by design. We handled this issue by applying different weighting schemes and reached robust answers. The strengths of our study are data availability and quality – recent, large, multi-year, representative samples with detailed outcomes-related information (i.e., fasting glucose, A1C and 2-hour glucose) and all of the necessary predictors measured. Although cross-sectional data are appropriate for undiagnosed or prevalent disease, the use of prospective data could provide additional useful insight or lessons (e.g., prediction of incident disease, if new risk score and/or new risk factor is needed; if more complicated model is warranted.)
In conclusion, the ADA and CDC scores performed well and comparably, and performance was robust to different data availability and deviations/modifications often entailed in practice. The same score may be used for DM and preDM. This direction may help active identification of preDM cases, which deserves to be a new focus of screening, by patients as well as healthcare providers in an efficient and seamless manner. While having preDM, one still has a chance to delay DM or even reverse the condition, during which some interventions have been shown to be effective and cost-effective. Despite limitations, self-assessment of DM and preDM risks has a potential to be the cheapest, easiest and safest way to learn about the risk and key risk factors, and to promote patient empowerment and patient-centered care.
Supplementary Material
Highlights.
We validated the well-known diabetes screening scores.
We showed that the same screening scores are well justified for the use of diabetes and prediabetes.
These screening scores could facilitate self-assessment and (pre-)screening of diabetes and prediabetes in community and clinical settings, and help health education for patients and lay persons.
Acknowledgments
Funding: H. Bang was partly supported by the National Center for Advancing Translational Sciences, National Institutes of Health, through grant number UL1 TR000002. We thank Dr. Lorena Garcia for motivating this manuscript and helpful discussions.
APPENDIX
A. Variables used in sensitivity analyses
For the sensitivity analyses, the following modifications were considered:
DM and preDM: We derived the outcomes based on fasting glucose only, A1C only, or 2-hour glucose only.
Hypertension: We used self-report, without blood pressure measurements.
Obesity: We added WC as a continuous covariate/predictor, in addition to BMI-based categories. Also, we derived the obesity categories by combining BMI and WC [2]. We tested the binary variable, overweight vs. under/normal weight based on self-assessment without anthropometric measures. The answer choice did not allow obese in the NHANES so this scenario represents a situation with underreporting of obesity, where no patients perceived they are obese [18]. Finally, we tested BMI vs. WC vs. WHtR when they were separately included as continuous predictor. Of note, another commonly considered measure, waist-hip ratio was not included in comparison because hip circumference was not available.
Physical activity: We derived two additional variables for physical activity. First, we derived a binary variable: Yes or No if >30 minutes of having recreational activities in a typical day. Second, we included ‘hours of sedentary activity in a typical day’.
Medical history: We simulated the scenarios where no pregnancy or family history data are available, which are common when these scores are used with administrative data or in some clinical settings [18].
B. Additional predictors tested in ancillary analyses
The following variables are not included in the original scores but we examined them as they are supported by the literature and could be useful for education or targeted screening [33 34]. We focused on modifiable factors among clinical, behavioral and dietary variables that patients are familiar with, in addition to race which is currently being used in various DM education programs. Thus, some of the variables may be used in future regression analyses or health education materials (e.g., information to be added to the back of the score card, in the follow-up step of the risk assessment) if educators or users wish.
Race: is categorized into 4 groups: Black (non-Hispanic); White (non-Hispanic); Hispanic; and Others.
Alcohol consumption: ‘average number of alcoholic drinks per day for the past 12 months’ was used.
Smoking: was analyzed in 2 manners: binary variable (current smoker vs. others) and continuous variable (average number of cigarettes per day for the past 30 days). Note that the NHANES did not use the same duration of time for alcohol and smoking.
Healthy diet: We used the answer (1: excellent to 5: poor) to “how healthy is your diet?”
Clinical and dietary: HDL, LDL, and total cholesterols, total sugar, and total fat were considered.
Footnotes
Conflicts of Interest: None
Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
References
- 1.Herman W, Smith P, Thompson T, et al. A new and simple questionnaire to identify people at increased risk for undiagnosed diabetes. Diabetes Care. 1995;18(3):382–87. doi: 10.2337/diacare.18.3.382. [DOI] [PubMed] [Google Scholar]
- 2.Bang H, Edwards A, Bomback A, et al. Development and validation of a patient self-assessment score for diabetes risk. Annals of Internal Medicine. 2009;151:755–83. doi: 10.1059/0003-4819-151-11-200912010-00005. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Rolka D, Narayan K, Thompson T, et al. Performance of recommended screening tests for undiagnosed diabetes and dysglycemia. Diabetes Care. 2001;24(11):1899–903. doi: 10.2337/diacare.24.11.1899. [DOI] [PubMed] [Google Scholar]
- 4.American Diabetes Association. Standards of medical care in diabetes--2014. Diabetes Care. 2014;37:S14–80. doi: 10.2337/dc14-S014. [DOI] [PubMed] [Google Scholar]
- 5.International Expert Committee report on the role of the A1C assay in the diagnosis of diabetes. Diabetes Care. 2009;32(7):1327–34. doi: 10.2337/dc09-9033. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Babey S, Wolstein J, Diamant A, et al. Prediabetes in California: Nearly half of California adults on path to diabetes. Los Angeles, CA: UCLA Center for Health Policy Research and California Center for Public Health Advocacy; 2016. [PubMed] [Google Scholar]
- 7.Xu Z, Qi X, Dahl A, et al. Waist-to-height ratio is the best indicator for undiagnosed type 2 diabetes. Diabetic Medicine. 2013;30(6):e201–7. doi: 10.1111/dme.12168. [DOI] [PubMed] [Google Scholar]
- 8.Schneider H, Friedrich N, Klotsche J, et al. The Predictive value of different measures of obesity for incident cardiovascular events and mortality. Journal of Clinical Endocrinology and Metabolism. 2010;95(4):1777–85. doi: 10.1210/jc.2009-1584. [DOI] [PubMed] [Google Scholar]
- 9.Chan D, Barrett H, Watts G. Recent studies of lipoprotein kinetics in the metabolic syndrome and related disorders. Current Opinion in Lipidology. 2006;17:28–36. doi: 10.1097/01.mol.0000199815.46720.ca. [DOI] [PubMed] [Google Scholar]
- 10.Diabetes Disparities Among Racial and Ethnic Minorities: Fact Sheet. Agency for Healthcare Research and Quality; Rockville, MD: [Accessed on April 29 2016]. 2001. http://archive.ahrq.gov/research/findings/factsheets/diabetes/diabdisp/diabdisp.html. [Google Scholar]
- 11.Cook N, Paynter N, Eaton C, et al. Comparison of the Framingham and Reynolds Risk scores for global cardiovascular risk prediction in the multiethnic Women’s Health Initiative. Circulation. 2012;125(13):1748–56. doi: 10.1161/CIRCULATIONAHA.111.075929. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.DeFilippis A, Young R, Carrubba C, et al. An analysis of calibration and discrimination among multiple cardiovascular risk scores in a modern multiethnic cohort. Annals of Internal Medicine. 2015;162:266–75. doi: 10.7326/M14-1281. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Zhang L, Zhang Z, Zhang Y, et al. Evaluation of Finnish Diabetes Risk Score in screening undiagnosed diabetes and prediabetes among U.S. adults by gender and race: NHANES 1999–2010. PloS ONE. 2014;9(5):e97865. doi: 10.1371/journal.pone.0097865. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Heikes K, Eddy D, Arondekar B, et al. Diabetes risk calculator: A simple tool for detecting undiagnosed diabetes and pre-diabetes. Diabetes Care. 2008;31:1040–45. doi: 10.2337/dc07-1150. [DOI] [PubMed] [Google Scholar]
- 15.Kahn H, Cheng Y, Thompson T, et al. Two risk-scoring systems for predicting incident diabetes mellitus in U.S. adults age 45 to 64 years. Annals of Internal Medicine. 2009;150:741–51. doi: 10.7326/0003-4819-150-11-200906020-00002. [DOI] [PubMed] [Google Scholar]
- 16.Lee Y-h, Bang H, Kim H, et al. A simple screening score for diabetes for the Korean population: Development, validation, and comparison with other scores. Diabetes Care. 2012;35:1723–30. doi: 10.2337/dc11-2347. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17. [Accessed on April 29 2016];Waist circumference and waist-hip ratio, report of a WHO expert consultation. 2008 http://apps.who.int/iris/bitstream/10665/44583/1/9789241501491_eng.pdf.
- 18.Shin J, Shortridge-Baggett L, Sachmechi I, et al. Screening for type 2 diabetes mellitus in patients with mental illness: application of a self-assessment score for diabetes mellitus risk. Psychiatry Research. 2014;220(3):1037–42. doi: 10.1016/j.psychres.2014.08.047. [DOI] [PubMed] [Google Scholar]
- 19.World Health Organization Expert Consultation. Appropriate body-mass index for Asian populations and its implications for policy and intervention strategies. Lancet 2004. 2004;363:157–63. doi: 10.1016/S0140-6736(03)15268-3. [DOI] [PubMed] [Google Scholar]
- 20.Schmidt M, Duncan B, Bang H, et al. Identifying individuals at high risk for diabetes: The Atherosclerosis Risk in Communities study. Diabetes Care. 2003;28:2013–18. doi: 10.2337/diacare.28.8.2013. [DOI] [PubMed] [Google Scholar]
- 21.Sun Q, van Dam R, Spiegelman D, et al. Comparison of dual-energy x-ray absorptiometric and anthropometric measures of adiposity in relation to adiposity-related biologic factors. American Journal of Epidemiology. 2010;172:1442–54. doi: 10.1093/aje/kwq306. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Bang H, Mazumdar M, Newman G, et al. Screening for kidney disease in vascular patients: SCreening for Occult REnal Disease (SCORED) experience. Nephrology Dialysis Transplantation. 2009;24:2452–57. doi: 10.1093/ndt/gfp124. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Kshirsagar A, Chiu Y-I, Bomback A, et al. A hypertension risk score for middle-aged and older adults. Journal of Clinical Hypertension. 2010;12:800–08. doi: 10.1111/j.1751-7176.2010.00343.x. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Wyatt J, Altman D. Prognostic models: clinically useful or quickly forgotten? BMJ. 1995;311:1539. [Google Scholar]
- 25.Gail M. Twenty-five years of breast cancer risk models and their applications. JNCI. 2015;107(5):djv042. doi: 10.1093/jnci/djv042. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Greenland S. The need for reorientation toward cost-effective prediction: comments on ‘Evaluating the added predictive ability of a new marker: From area under the ROC curve to reclassification and beyond’. Statistics in Medicine. 2008;27(2):199–206. doi: 10.1002/sim.2995. [DOI] [PubMed] [Google Scholar]
- 27.Lee Y, Bang H, Kim D. How to establish clinical prediction models. Endocrinology and Metabolism. 2016;31(1):38–44. doi: 10.3803/EnM.2016.31.1.38. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Thamer M, Kaufman J, Zhang Y, et al. Predicting early death among elderly dialysis patients: development and validation of a risk score to assist shared decision making for dialysis initiation. American Journal of Kidney Disease. 2015;66(6):1024–32. doi: 10.1053/j.ajkd.2015.05.014. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Herman W, Smith P, Thompson T, et al. Response to Knudson et al. Diabetes Care. 1998;21(6):1030–31. [Google Scholar]
- 30.Dawes R, Faust D, Meehl P. Clinical versus actuarial judgement. Science. 1989;243:1668–74. doi: 10.1126/science.2648573. [DOI] [PubMed] [Google Scholar]
- 31.Kim G, Lee Y-h, Park Y, et al. Use of a diabetes self-assessment score to predict nonalcoholic fatty liver disease and nonalcoholic steatohepatitis. Medicine. 2015;94(27):e1103. doi: 10.1097/MD.0000000000001103. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Liao L, Mark D. Clinical prediction models: Are we building better mousetraps? Journal of the American College of Cardiology. 2003;42(5):851–53. doi: 10.1016/s0735-1097(03)00836-2. [DOI] [PubMed] [Google Scholar]
- 33.Bardenheier B, McKeever K, Caspersen C, et al. A novel use of structural equation models to examine factors associated with prediabetes among adults aged 50 years and older. Diabetes Care. 2013;2013(36):2655–62. doi: 10.2337/dc12-2608. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34.Buijsse B, Simmons R, Griffin S, et al. Risk assessment tools for identifying individuals at risk of developing type 2 diabetes. Epidemiologic Reviews. 2011;33(1):46–62. doi: 10.1093/epirev/mxq019. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.