Using data from a cross-sectional survey in a Japanese community, we demonstrated that the optimal thresholds for detecting prevalent DR from ROC analyses were 6.5 mmol/l for FPG, 11.5 mmol/l for 2-hour PG, 6.1% (43 mmol/mol) for HbA1c, 17.0% for GA, and 12.1 μg/mL for 1,5-AG. These results were in accordance with those from the prevalence analysis of DR by decile levels of these measures of glycemia. These findings suggest that the FPG and HbA1c thresholds for diagnosing diabetes in the Japanese population are lower than the current diagnostic criterion, while the 2-hour PG threshold is approximately 11.1 mmol/l, which is comparable to the diagnostic criterion. To our knowledge, the present study is the first report to determine the GA and 1,5-AG thresholds for the diagnosis of diabetes using the prevalence of DR. Furthermore, 2-hour PG had higher sensitivity and larger AUC than other glycemic measures, whereas the AUCs for FPG, HbA1c, GA, and 1,5-AG were not significantly different. These findings indicate that 2-hour PG has the highest discriminative ability, and measurements of FPG, HbA1c, GA, and 1,5-AG are similar in their ability.
The HbA1c thresholds for identifying presence of DR have varied among prior epidemiological studies of Asian populations, ranging from 6.1% (43 mmol/mol) to 7.0% (53 mmol/mol). In a study in a Singapore population, the optimal HbA1c threshold for DR was 6.6-7.0% (49-53 mmol/mol) . A subanalysis of the DETECT-2, which included three Asian studies in India, Singapore, and Japan, showed that an HbA1c of 6.4% (46 mmol/mol) was the optimal threshold . Similar findings were observed in a Chinese population study (6.4% [46 mmol/mol]) . On the other hand, in the present study, the prevalence of DR increased precipitously when HbA1c levels exceeded 5.9-6.2% (41-44 mmol/mol), and the optimal threshold of HbA1c using the ROC analysis was 6.1% (43 mmol/mol). Importantly, this threshold was identical with that from our previous study in 1998 (NGSP: 6.1% [43 mmol/mol]; JDS: 5.7%) . Furthermore, in another epidemiological study in a Japanese population, the ROC analysis indicated that the highest precision for DR was observed at an HbA1c value of 6.2% (44 mmol/mol) . These findings suggest that the HbA1c threshold in the Japanese population was lower than that of other Asians and also lower than the diagnostic criterion of 6.5% (48 mmol/mol). Although the reason for this difference is unclear, the influence of race and ethnicity on HbA1c levels may contribute to this phenomenon. Some epidemiological studies have shown that South Asians, Hispanics, and Blacks had higher HbA1c levels than non-Hispanic whites, independent of glucose [39, 40]. In our subjects, the mean of HbA1c levels (5.5% [37 mmol/mol]) was lower than those in other Asian population studies (6.0-6.5% [42-48 mmol/mol]) [15–17]. Thus, it is speculated that there are racial and ethnic differences in HbA1c levels even among Asians, and this may be the reason for the lower threshold in our subjects. In addition, the aldehyde dehydrogenase 2*2 (ALDH2*2) allele, which is more common in East Asians than in other ethnic groups, has been identified as a genetic risk factor for incident DR in Japanese subjects with diabetes , and thus, such a genetic difference in susceptibility to DR might also affect the HbA1c levels associated with incident DR.
The use of HbA1c measurement to diagnose diabetes remains somewhat controversial . Recent epidemiological studies have revealed that HbA1c measurement alone was less sensitive for detecting subjects with diabetes compared to the OGTT [43, 44]. However, in our study, the AUCs for HbA1c and FPG were not significantly different. This finding indicates that the discriminative ability of HbA1c for diagnosing diabetes was comparable to that of FPG. Furthermore, HbA1c measurement can be done without fasting or timed samples, and thus it would be suitable for mass screening in general practice. This advantage has implications for the early identification and treatment of undiagnosed diabetes. For these reasons, HbA1c measurement may be an appropriate tool for detecting undiagnosed diabetes. In addition, some clinical and population-based studies, including our previous study, have shown that elevated HbA1c levels were independently associated with cardiovascular disease [45, 46], suggesting that HbA1c is also useful as a predictor of macrovascular complications. Therefore, the use of HbA1c to diagnose diabetes will help to prevent both micro- and macrovascular complications of diabetes, which are increasingly recognized as a global health priority.
In the present analysis, although the prevalence of DR was quite small for GA below 16.2-17.5%, it began to rise sharply above these levels, and the optimal threshold of GA using the ROC analysis was 17.0%. Several studies have examined the use of GA levels for detecting diabetes or glucose intolerance, as defined by glucose levels. In a Japanese population study, the ROC analysis for detecting diabetes identified the GA threshold as 15.5% , while another study of Japanese subjects reported a GA level of 17.0% as the lower limit of glucose intolerance . A similar threshold of GA was obtained in a Chinese population study (17.1%) . The thresholds in these studies were in good agreement with our findings. Together with those of other studies, our findings suggest that the optimal GA threshold for diagnosing diabetes is likely to be 17.0%.
There have been a few studies evaluating the optimal threshold of 1,5-AG for identifying individuals with diabetes, as defined by a OGTT. A Japanese population study showed that 14.0 μg/mL was the best value for detecting subjects with diabetes . Similar findings were observed among Japanese male workers (14.2 μg/mL) . In a Chinese study, the mean of 1,5-AG levels was 15.0 μg/mL in subjects with newly diagnosed diabetes and 11.8 μg/mL in subjects with known diabetes . However, no study showed an optimal threshold of 1,5-AG using the presence of DR. The present study revealed that the steepest increment in the prevalence of DR occurred when the 1,5-AG levels fell below 9.6-13.5 μg/mL, and that 12.1 μg/mL was the optimal 1,5-AG threshold in the ROC analysis. Further epidemiological studies are needed to verify our findings.
The FPG of 7.0 mmol/l and the 2-hour PG of 11.1 mmol/l for diagnosing diabetes with the current criterion were also derived mainly from studies in Western populations . In our study, the optimal threshold for detecting prevalent DR was 6.5 mmol/l for FPG, and 11.5 mmol/l for 2-hour PG. The 2-hour PG threshold was compatible with that from our previous report (11.1 mmol/l)  and another study of a Japanese population (11.0 mmol/l) . Meanwhile, other Asian population studies have reported that the optimal FPG threshold for DR was 7.0 mmol/l in a Japanese population , and 7.2 mmol/l in a Chinese population . These findings are inconsistent with ours. However, a recent meta-analysis in Asian and Western populations evaluated the relationship of glucose levels with DR and concluded that the FPG threshold for diagnosing diabetes was 6.5 mmol/l . Furthermore, our prior studies showed that the threshold of FPG for DR was 6.4 mmol/l  and that the FPG threshold corresponding to a 2-h PG of 11.1 mmol/l was 6.2 mmol/l . These results were very similar to those of the present study. Taken together, these findings imply that, in a Japanese population, the threshold of FPG for diabetes is lower than the diagnostic criterion of 7.0 mmol/l, while the threshold of 2-hour PG is 11.1 mmol/l, which is in accord with the diagnostic criterion.
The present study showed that among the five glycemic measures, 2-hour PG had not only the highest sensitivity but also the largest AUC to identify the presence of DR. These results suggest that the performance and discriminative ability of 2-hour PG for diagnosing diabetes were superior to those of other glycemic measures. Oxidative stress is known to be one of the crucial contributors in the pathogenesis of DR . It has also been reported that acute hyperglycemia had a more specific triggering effect on oxidative stress than chronic sustained hyperglycemia [49, 50]. Thus, 2-hour PG values can be considered a better marker of oxidative stress levels arising from acute hyperglycemia than FPG, HbA1c, GA, and 1,5-AG values. Furthermore, a prospective study demonstrated that postprandial plasma glucose was a stronger predictor of the progression of DR than HbA1c in Japanese subjects with diabetes . Taken together, these findings imply that 2-hour PG levels may be more strongly associated with DR than other glycemic measures. This may explain why 2-hour PG has a high diagnostic accuracy for DR. On the other hand, the AUCs for GA and 1,5-AG did not significantly differ from those for FPG and HbA1c, suggesting that GA and 1,5-AG are acceptable alternatives for the diagnosis of diabetes, and these two measures may be particularly useful for individuals with anemia, renal disease or hemoglobinopathy, for whom interpretation of HbA1c values is problematic. However, the 1,5-AG levels had smaller AUC with lower sensitivity than other glycemic measures. One possible explanation for this phenomenon may be that 1,5-AG levels reflect the degree of glycosuria rather than glucose levels , while other glycemic measures directly indicate the degree of hyperglycemia. In addition, it has been reported that 1,5-AG levels were influenced by individual difference in their renal thresholds for glucose . These facts might be the reason for the relatively low discriminative ability of 1,5-AG in our study.
The strengths of our study include the population-based design, high participation rate, and availability of data to evaluate the five glycemic measures. In addition, it is noteworthy that the FPG, 2-hour PG, and HbA1c thresholds in the present study were nearly the same as those from our previous study , suggesting the high reproducibility of the results in our population. However, some limitations should also be mentioned. First, our analyses included subjects with antidiabetic medications. Hypoglycemic medications could have affected the levels of glycemia. The optimal thresholds remained substantially unchanged, except for GA, after excluding subjects who received hypoglycemic medications (FPG: 6.3 mmol/l; 2-hour PG: 11.5 mmol/l; HbA1c: 6.2% [44 mmol/mol]; GA: 20.5%; and 1,5-AG: 12.1 μg/mL). However, the precision of this finding may be limited, because of the small number of cases of DR among those not taking hypoglycemic medications. Second, the values of HbA1c were not measured by high-performance liquid chromatography (HPLC) as used in the Diabetes Control and Complications Trial, although the method and reagent used to measure HbA1c in this study have since been NGSP-certified. It would be preferable to measure HbA1c by HPLC to make the results of our study more comparable to those of other studies. Third, the GA and 1,5-AG levels were measured in serum conserved at -80°C for 5 years. However, the stability of GA and 1,5-AG measurements in frozen stored serum sample was preserved [53, 54]. Fourth, this study is a cross-sectional design, which might have affected the threshold values of glycemic measures. Diagnostic thresholds would ideally be derived from prospective studies that examine the relationship between measures of glycemia and incident microvascular complications. Lastly, the influence of factors that may affect HbA1c levels, such as anemia, renal failure, and hemoglobinopathy should be considered. We performed sensitivity analyses excluding subjects with anemia or renal failure, and the optimal threshold of HbA1c remained unchanged (6.1% [43 mmol/mol]). Furthermore, the prevalence of hemoglobinopathy in Japan was reported to be very low (0.04%) . Therefore, the influence of this limitation would have been small.