The Performance of a Risk Score as a Screening Test for Undiagnosed Hyperglycemia in Ethnic Minority Groups

Data from the 1999 Health Survey for England

  1. Annemieke M.W. Spijkerman, PHD1,
  2. Matthew F. Yuyun, MB, PHD2,
  3. Simon J. Griffin, MD2,
  4. Jacqueline M. Dekker, MB1,
  5. Giel Nijpels, MB1 and
  6. Nicholas J. Wareham, MB, PHD2
  1. 1Institute for Research in Extramural Medicine, VU University Medical Center, Amsterdam, the Netherlands
  2. 2Department of Public Health and Primary Care, University of Cambridge, Cambridge, U.K.
  1. Address correspondence and reprint requests to Dr. Nicholas Wareham, Department of Public Health and Primary Care, University of Cambridge, Institute of Public Health, University Forvie Site, Robinson Way, Cambridge CB2 2 SR, U.K. E-mail: njw1004{at}medschl.cam.ac.uk

Abstract

OBJECTIVE—To assess the performance of the Cambridge Risk Score (CRS) to predict undiagnosed hyperglycemia in Caribbean and South Asian people living in the U.K.

RESEARCH DESIGN AND METHODS—The CRS uses routinely available data from primary care records to identify people at high risk for undiagnosed type 2 diabetes. The sensitivity, specificity, and area under the receiver operator characteristic (ROC) curve for the CRS cut point of 0.199 were 77, 72, and 80% (95% CI 68–91), respectively. The risk score was calculated for 248 Caribbean and 555 South Asian participants aged 40–75 years in the 1999 Health Survey for England. Undiagnosed hyperglycemia was considered present if fasting plasma glucose was ≥7.0 mmol/l or HbA1c was ≥6.5%. Sensitivity, specificity, and predictive values were calculated for various cut points of the risk score, and ROC curves were constructed.

RESULTS—The area under the ROC curve was 67% (59–76) and 72% (67–78) for Caribbeans and South Asians, respectively. The optimal cut point in Caribbean participants was 0.236, sensitivity was 63% (46–77), and specificity was 63% (56–69). In the South Asian population, the optimal cut point was and 0.127, sensitivity was 69% (60–78), and specificity was 64% (60–69).

CONCLUSIONS—The CRS, using routinely available data, can be used in a strategy to detect undiagnosed hyperglycemia in Caribbean and South Asian populations. The existence of ethnic group–specific cut points must be further established in future studies.

The high prevalence of undiagnosed diabetes (13) and the demonstration of evidence of diabetic complications at the time of diagnosis (4) have led to recommendations for screening for type 2 diabetes by the American Diabetes Association (ADA) and Diabetes U.K. (5,6). One possible way of screening for type 2 diabetes is with simple risk scores based on data that are routinely available in primary care (7,8). Risk scores and questionnaires for type 2 diabetes have been developed and tested in mainly Caucasian populations (710), but it is uncertain whether these instruments could be used to identify undiagnosed diabetes in non-Caucasian populations as well. The ADA questionnaire is the only screening tool including a question about ethnicity (11). To our knowledge, no risk score for type 2 diabetes has been developed specifically for non-Caucasian populations. The demonstration of the validity of existing scores in specific ethnic groups or the development of ethnic-specific risk scores would be important because the prevalence of type 2 diabetes in individuals of African-Caribbean and South Asian origin is high, ranging from 14.6 to 17% in Caribbeans (12,13) and from 20 to 25.4% in South Asians (1315); ∼40% of cases are undiagnosed (16).

We have previously developed and tested a risk score that uses routinely available data from primary care records to identify individuals at high risk for undiagnosed type 2 diabetes. In the U.K., virtually all individuals in the population are registered with a primary care practice, making these primary care records a population-based sampling frame. The risk score seemed to be effective at identifying individuals with previously undiagnosed diabetes or hyperglycemia in population groups that were predominantly Caucasian (8,17).

In this study, we tested the performance of the risk score in detecting undiagnosed hyperglycemia, as measured by fasting glucose or HbA1c, in the ethnic minority sample of the 1999 Health Survey for England (HSE), a population-based study.

RESEARCH DESIGN AND METHODS

We used data from the ninth version (1999) of the HSE, a series of annual surveys designed to monitor trends in health status in the U.K., based on data from nationally representative samples (18). The 1999 HSE was designed to augment existing research on the health of ethnic minority groups. To enhance the number of people from ethnic minority groups, the size of the general population sample was restricted and the 1999 survey included a large-scale representative sample of minority ethnic adults and children throughout the country. The sample of ethnic minority groups included people from the most populous minority ethnic groups in the U.K.: individuals from the Caribbean, India, Pakistan, Bangladesh, China, and Ireland. A sample of 64,000 addresses were selected from 340 postal sectors in the U.K., and at each address, a maximum of four adults and three children were included in the survey.

In our study, we focused on Caribbean people and people from India, Pakistan, and Bangladesh without previously diagnosed type 2 diabetes and between the ages of 40 and 75 years. We did not include the Chinese sample because of the small number of Chinese participants with complete data on risk score, glucose, and HbA1c. Because of small numbers in the ethnic groups from the Indian subcontinent, we decided to create one group of South Asians including the participants from India, Pakistan, and Bangladesh (n = 1,174 South Asians, n = 488 Caribbeans). Participants with missing data on the risk score were excluded (82 Caribbean and 178 South Asian participants). The population with a valid risk score was further restricted to individuals with data on glucose and/or HbA1c. Consequently, current analyses were performed using data from 248 Caribbean and 555 South Asian participants.

The HSE consisted of an interview, a nurse visit, random blood sampling, and fasting blood sampling. Informed consent was given for each part separately. The HSE was approved by the North Thames Multi-Centre Research Ethics Committee and all local ethics committees in England.

Measurements

The methods for the HSE have been reported elsewhere (18). In brief, prescribed drugs and smoking habits were recorded during the interview and data on social class and physical activity were obtained. Height, weight, waist circumference, and hip circumference were measured without shoes and with the participants wearing only light clothing. Blood pressure was measured with a Dinamap 1800 automated blood pressure monitor after a 5-min rest. The mean of three measurements was used in the analysis. Blood pressure measurements were available for 197 Caribbean and 484 South Asian participants. Fasting plasma glucose was determined by the specific glucose oxidase mediated peroxidase/4-aminoantipyrine method (19). Measurement of HbA1c was performed using the Tosoh HLC-723 (GHbV) A1c2.2 analyzer (20), a Diabetes Control and Complications Trial (DCCT) traceable method. The mean intra-assay coefficients of variation were 3.72 and 2.50% for the lower and upper end of the range, respectively (18).

Because fasting plasma glucose samples were not available for all participants (n = 183 and n = 394 for Caribbeans and South Asians, respectively), undiagnosed hyperglycemia was considered present if fasting plasma glucose was ≥7.0 mmol/l (21) or HbA1c was ≥6.5%. HbA1c has not been included in the diagnostic criteria of either the ADA or the World Health Organization for reasons of standardization and global availability. Fasting plasma glucose, 2-h glucose, and HbA1c have been shown to be equally predictive of future retinopathy (22). In addition, HbA1c has been shown previously to be associated with microvascular complications of diabetes (23) and all-cause and cardiovascular mortality (24). As such, HbA1c is a reasonable measure of undiagnosed hyperglycemia. Cut points for HbA1c in the literature range from 6.1 to 7.8% (22,23,25,26), depending on the reference method used (fasting plasma glucose, 2-h glucose, or future risk of retinopathy). In this study, a cut point of 6.5% was chosen to represent undiagnosed hyperglycemia. An HbA1c value of 6.1% showed good test performance in African-American participants of the National Health and Nutrition Examination Survey III (NHANES III) (25). Moreover, a previous study of the Cambridge Risk Score (CRS) in Caucasians showed that the lowering of the diagnostic threshold of HbA1c from 7.0 to 6.5% did not materially change the performance of the CRS as measured by the area under the receiver operator characteristic (ROC) curve (17). Equivalence of the fasting glucose and HbA1c diagnostic threshold was not our objective, because we considered the two measures to be separate indicators of undiagnosed hyperglycemia.

Screening instrument: CRS

The risk score was developed to determine the probability of having undiagnosed type 2 diabetes with a logistic regression model using data that would be routinely available in primary care records, as has been described previously (8). The CRS is based on the sum of the score for each of the variables included in the model, i.e., age, sex, prescribed antihypertensive medication and corticosteroids, family history of diabetes, BMI, and smoking status (appendix). The total score (between 0 and 1) gives the probability of having undiagnosed diabetes; the higher the score, the higher the chance of having diabetes. The CRS was developed and evaluated against the results of the oral glucose tolerance test (OGTT), the gold standard for type 2 diabetes. The sensitivity, specificity, and area under the ROC curve for the previously published cut point of 0.199 (8) were 77, 72, and 80% (95% CI 68–91), respectively. The CRS is used as a first step in a targeted screening program for diabetes to select individuals at high risk for undiagnosed diabetes to undergo diagnostic testing. In the present study, we calculated the CRS with data that were already collected for the 1999 HSE. Data on family history of diabetes were not available, and the CRS was calculated with data on age, sex, prescribed antihypertensive medication and corticosteroids, BMI, and smoking status.

Statistical analyses

Biochemical and anthropometric characteristics were compared between hyperglycemic and normoglycemic participants using Student’s t test comparison for normally distributed variables. Differences in categorical variables were assessed with χ2 tests. Sensitivity, specificity, positive and negative predictive value, and likelihood ratio (LR) and 95% CI were calculated for various cut points. The sensitivity of the CRS is the proportion of people with undiagnosed type 2 diabetes who are correctly identified as such by the CRS (percentage true positives). The specificity of the CRS is the proportion of people without undiagnosed type 2 diabetes who are correctly identified as such by the CRS (percentage true negatives). The positive predictive value of the CRS is the proportion of people with a positive result on the CRS who indeed have undiagnosed type 2 diabetes. The negative predictive value of the CRS is the proportion of people with a negative result on the CRS who indeed do not have undiagnosed type 2 diabetes. The LR of a positive test (LR+), or the odds of being affected given a positive result, is the ratio of the percentage true positives to the percentage false positives. The LR of a negative test (LR–) is the ratio of the proportion of false negatives to the proportion of true negatives. The LR is used to calculate the (post-test) odds of a disease after the (screening) test took place. ROC curves were plotted, and the area under the curve and the 95% CI were estimated (27). The larger the area under the curve, the better performance of the screening test. CIs for the performance characteristics of the risk score were calculated (28). A P value of <0.05 was considered statistically significant, based on two-sided testing. SPSS for Windows version 10 (SPSS, Chicago, IL) was used for all analyses.

RESULTS

The Caribbean sample included 248 participants, and there were 555 individuals in the South Asian sample. The clinical characteristics of the study groups are shown in Table 1. Undiagnosed hyperglycemia was common, occurring in 40 (16.1%) of the Caribbean and 102 (18.4%) of the South Asian participants.

The characteristics of the Caribbean and South Asian subjects with undiagnosed hyperglycemia and normoglycemia are shown in Table 2. Caribbean participants with undiagnosed hyperglycemia were significantly older, more obese, and had significantly higher systolic blood pressure than normoglycemic Caribbean participants.

In undiagnosed hyperglycemic individuals from South Asia, systolic blood pressure was significantly higher and use of antihypertensive medication was more frequently reported than in normoglycemic South Asian participants. In addition, hyperglycemic South Asians were significantly older and were more obese. Their obesity was also more likely to be centrally deposited, as demonstrated by a higher waist-to-hip ratio (WHR). In both ethnic groups, glucose and HbA1c were significantly higher, by definition, in undiagnosed hyperglycemic participants compared with normoglycemic individuals.

The test characteristics of the CRS for the prediction of undiagnosed hyperglycemia for various cut points are shown in Table 3. The optimal cut point (high sensitivity with comparable high specificity) for the CRS was different in the two ethnic groups. In Caribbean people, the optimal cut point was 0.236; sensitivity was 63% (95% CI 46–77), specificity was 63% (56–69), positive predictive value was 24% (16–33), and negative predictive value was 90% (85–95). In the South Asian population, the optimal cut point was 0.127; sensitivity was 69% (60–78), specificity was 64% (60–69), positive predictive value was 30% (24–36), and negative predictive value was 90% (86–93).

Figure 1 shows the ROC curves of the risk score for undiagnosed hyperglycemia. The areas under the curve were lower in Caribbean (67% [59–76]) and South Asian (72% [67–78]) individuals compared with the area under the curve in the original test population (80% [68–91]).

CONCLUSIONS

This is the first study to test the performance of a simple risk score using routinely collected data as a screening tool for undiagnosed hyperglycemia in ethnic minority groups. Although the performance of the risk score was not as good as in Caucasian people, the area under the ROC curve indicates that the risk score could have a role in strategies to identify people with undiagnosed hyperglycemia in Caribbean and South Asian populations. For reasons of simplicity, one might want to use a single cut point for all populations, but this would not be logical if it resulted in poor sensitivity or specificity. The data from this study suggest that different cut points for the CRS should be used in different ethnic groups.

The ethnic sample of the HSE was designed to oversample Caribbean and South Asian people so that they constituted a larger part of the total study population in the survey than they do in the general U.K. population. Response rates for fasting blood samples were lower than reported in other studies measuring fasting glucose (12,29). Informed consent for various parts of the study was given separately and consecutively, which might have negatively influenced the participation rate for the fasting blood sample because it was last. In South Asians, participants with fasting glucose measurements were significantly more likely to be in a higher social class and more physically active than individuals without glucose measurements. Participants with HbA1c measurements were significantly younger in both ethnic groups. In addition, in the South Asian group, the participants with HbA1c measurements reported a higher social class and higher physical activity level and were more likely to be men than the South Asian participants without HbA1c measurements. This selection of a group of relatively healthy people might have resulted in an underestimation of the performance of the CRS (lower positive predictive value because of the lower prevalence of undiagnosed hyperglycemia). A sample without this selection might have improved the already acceptable test performance.

Population-based OGTT or universal screening studies have been performed in Caribbean and South Asian populations (12,14,16,30) to establish the prevalence of type 2 diabetes in these ethnic groups, but few reports exist about the effectiveness of targeted screening for type 2 diabetes. Davies et al. (31) performed a study in South Asians to determine the effectiveness of a targeted screening program that consisted of self-testing for postprandial glycosuria as a screening test followed by OGTT. The reported low response rate for the glycosuria test (34%) (31) diminished the overall effectiveness of this screening program. In contrast, a screening program with the CRS as a screening test would not suffer from low response rates because the CRS can be calculated with data routinely available in the database of the primary care practice without having to contact the patient. After the calculation of the risk score, diagnostic testing would be restricted to the people at high risk, and this restriction would enhance the efficiency of the CRS. In addition, using the CRS, there would be no need for translations of questions or instruction sheets in the case of non–English-speaking people from ethnic minorities.

Because glucose measurements were not available for all participants in the 1999 HSE, HbA1c was used in addition to glucose as an indicator of undiagnosed hyperglycemia (32). Although HbA1c is associated with the development of microvascular and macrovascular complications, it has only recently been proposed as a diagnostic test for type 2 diabetes (25,33). Undoubtedly, there will be some differences between the individuals diagnosed by fasting glucose and by HbA1c, just as there is between those diagnosed by the 2-h value of the OGTT and the fasting plasma glucose value (34). However, it is unlikely that this variation will affect the performance of the CRS in the present study.

In this study, no data were available on family history of diabetes, one of the variables included in the CRS. We elected to use the CRS total score and the previously published cut points regardless of the lack of information on family history. A previous study in Caucasians showed that the risk score performed well, even in the absence of information on family history of diabetes (17). However, it is uncertain whether this result can be generalized to non-Caucasian population groups, because family history is more prevalent in Asian populations than in Caribbean and European people (35,36). Inclusion of data on (positive) family history might improve the performance of the CRS in non-Caucasians.

Higher age, higher levels of obesity, and hypertension seem to place individuals in a high-risk category for type 2 diabetes, regardless of ethnic group. The CRS includes all these risk factors, and it was plausible to expect an acceptable performance in ethnic minority groups. However, the distribution of these risk factors might differ between ethnic groups and may affect test performance. The optimal cut point was higher in Caribbeans and might be due to the higher prevalence of hypertension reported in Caribbean people (13,37). As a result, hypertension and use of antihypertensive medication might not be as distinctive a feature for diabetes in a Caribbean population as it is in Caucasians. The lower cut point for the South Asian participants might be ex-plained, in part, by the fact that the mean age of onset of diabetes is 5 years younger in South Asians than in Caucasians (16,35). Therefore, age does not contribute as much to the risk score as it would in a Caucasian population. In addition, the prevalence of current smoking was quite low, again resulting in a lower risk score. Moreover, in a risk score for undiagnosed hyperglycemia for South Asian people, inclusion of WHR instead of BMI might be more effective because central obesity is highly prevalent in South Asians (29,35) and seems to be more strongly associated with glucose intolerance than BMI (38). However, data on WHR would not be routinely available in primary care records and would decrease the utility of the risk score; therefore, BMI should remain to be used in the risk score for ethnic minorities.

In conclusion, our study showed that the CRS, using routinely available data, could be used as a first test in a screening strategy to identify individuals with undiagnosed hyperglycemia in a Caribbean and South Asian population. Further studies are needed to establish the ethnic-specific cut points observed in our study.

APPENDIX

The Cambridge Risk Score (8)

Figure 1—

ROC curve of the CRS for undiagnosed hyperglycemia in Caribbean (A) and South Asian (B) participants of the HSE.

Table 1—

Baseline characteristics according to ethnicity

Table 2—

Characteristics of Caribbean and South Asian individuals with undiagnosed hyperglycemia versus normoglycemia

Table 3—

Performance of the risk score in predicting undiagnosed hyperglycemia in Caribbean and South Asian participants of the HSE

Acknowledgments

A.S. was supported by a Travel Fellowship of the European Foundation for the Study of Diabetes and Merck & Co Inc. (EFSD/MSD Fellowship).

Footnotes

  • A table elsewhere in this issue shows conventional and Système International (SI) units and conversion factors for many substances.

    • Accepted September 15, 2003.
    • Received March 7, 2003.

References

| Table of Contents