Comparison Of The Optimal Cut-off Point Of HbA1c For Diagnosing Diabetes In Different Ethnic Populations

Part 1 Comparison of the optimal cut-off points of HbA1c for diagnosing diabetes in different ethnic populationsBackground and Aim:China is a multi-ethnic country with a large population and territory.Due to the ethnic and cultural differences,a tailored hemoglobin A1c（HbA1c）diagnostic criteria for diabetes,based on the local epidemiological surveys,is needed.However,there is still a lack of evidence for the HbA1c threshold to diagnose diabetes in China,especially based on a multi-ethnic database.In the previous study,we haven’t found the optimal cut-off point of HbA1c for Chinese population.This study will explore the optimal HbA1c threshold for diagnosing diabetes using the oral glucose tolerance test（OGTT）as a reference standard,and compare its efficiency based on the specific diabetic retinopathy（DR）.Further subgroup analyses will be based on different ethnicities,ages,and genders.Methods:This population-based cross-sectional study involved a randomized cluster sample of 20-70-year-old residents from 8 provinces in 6 ethnic groups（the Han,the Zhuang,the Dai,the Uyghur,the Korean and the Kazakh）.The OGTT and two retinal fundi（one disc and one macula centered 45°）images were conducted and collected for each participant.Other laboratory data including HbA1c,renal function and lipid metabolism parameters were also collected.Demographic information and physical examination results including height,weight,waist circumference,blood pressure and heart rate of each participant were obtained.OGTT was evaluated based on the 1999 World Health Organization（WHO）standard.The Early Treatment Diabetic Retinopathy Study（ETDRS）scale was used to diagnose DR,and the specific DR was defined by ETDRS≥20（very mild non-proliferative diabetic retinopathy）,ETDRS≥31（mild non-proliferative diabetic retinopathy）and ETDRS≥41（moderate non-proliferative diabetic retinopathy）,respectively.Statistical analysis was as follows:1)The overall data was divided into the training set and the validation set with a ratio of2:1 with the random sampling method.2)Using the OGTT results and ETDRS≥31 as the reference standard respectively,we explored the optimal HbA1c threshold with the receiver operating curve（ROC）based on the Youden Index（YI）.The same method was used for subgroup analysis,and Hanley&Mc Neil method was used to compare different ROC curves.3)The line graph between different HbA1c levels and the DR prevalence indicated the change point of HbA1c,varying from different ETDRS standards.Logistic regression was used to analyze the relationship between HbA1c and the risk factors of DR.4)A multi-factor linear model was used to analyze the factors that affect HbA1c levels.Logistic regression was used to analyze the factors affecting the false negative or positive population.5)The associated diagnosis model was achieved by logistic regression and ROC.Result:（1）Baseline resultsA total of 16,969 participants were included in the statistical analysis,with 11,312 in the training set,and 5,657 in the validation set.Based on this study,the prevalence of diabetes and prediabetes was 13.53%（2296）and 24.60%（4175）,respectively,and 61.87%（10498）of participants had normal glucose regulation.（2）The optimal HbA1c cut-off point based on OGTT results1.The area under the curve（AUC）of HbA1c was 0.886（95%CI,0.875-0.896）,and the optimal cut-off point was an HbA1c of≥6.0%.Based on the threshold,the sensitivity was73.14%（95%CI,70.9%-75.3%）,the specificity was 90.45%（95%CI,89.8%-91.0%）,the positive predictive value was 54.8%（95%CI,52.6%-56.9%）,and the negative predictive value was 95.5%（95%CI,95.1%-95.9%）.2.When the optimal cut-off point（HbA1c≥6.0%）was used:the factors including age,high LDL-C level,the Dai,or the Uyghur were associated with a decreased risk of false-negative diagnosis.Age,being overweight or obese,high LDL-C,the Dai,the Zhuang,or the Uyghur were associated with an increased risk of false-positive diagnosis.While high blood pressure,being thin,and the Kazakh were associated with a reduced risk of false-positive diagnosis.3.When categorized by ethnicity,the optimal HbA1c cut-off point of each ethnic group was close to 6.0%,except for the Dai（6.3%）.When categorized by age,the optimal HbA1c cut-off point increased with aging.The cut-off point was 6.1%for senior group（age>52）,5.9%for middle-aged group（36<age≤52）,and 5.7%for youth group（age≤36）.The optimal HbA1c cut-off points for different gender groups and different BMI groups were all6.0%.（3）The optimal HbA1c cut-off point based on specific DR1.The prevalence of specific DR had a noticeable change point when the HbA1c was about 6.3%.After adjusting by age,gender,BMI,and ethnicity,when the HbA1c level was above 6.3%,the risk of specific DR was significantly higher than the 3.0%-5.7%interval[ETDRS≥31:OR 32.01（17.79,57.61）,ETDRS≥20:OR 6.71（4.91,9.170）].2.Using ETDRS≥31 as the reference standard,the ROC curve was achieved.The AUC of HbA1c was 0.862（95%CI,0.810 to 0.913）,and the optimal cut-off point was HbA1c≥6.2%.Using the optimal cut-off point to diagnose diabetes,the sensitivity was 76.19%（95%CI,65.7%-84.8%）,the specificity was 88.98%（95%CI,88.1%-89.8%）,the positive predictive value was 6.1%（95%CI,5.1%-8.2%）,and the negative predictive value was99.6%（95%CI,99.6%-99.8%）.3.When the optimal diagnosis cut-off point（HbA1c≥6.2%）used,the factors including age,being a woman,being overweight or obese,high LDL-C,the Dai,or the Uygur showed an increased risk of false positive diagnosis.The Zhuang was associated with a reduced risk of false positive diagnosis.4.When categorized by ethnicity,the optimal HbA1c cut-off point was 6.1%for the Han,5.9%for the Dai,6.8%for the Uyghur,and 6.9%for the Korean.When categorized by age,the cut-off point was 6.8%for the senior group,and 6.2%for the middle-aged group.The HbA1c cut-off points varied with gender（Male:6.8%vs Female:6.1%）and BMI group（Normal:6.0%vs overweight and obesity:6.2%）.（4）HbA1c combined with fasting plasma glucose（FPG）for diagnosing diabetes1.HbA1c≥6.3%and/or FPG≥7.0mmol/L:The AUC was 0.788（95%CI,0.750-0.827）,which was not significantly different from AUC_OGTT 0.794（95%CI,0.786 to 0.802）（P=0.7085）.2.HbA1c+FPG:the optimal cut-off point was 0.00085+1.32*FPG+1.17*HbA1c>4.97.The AUC was 0.864（95%CI,0.854-0.874）,the sensitivity was 79.73%（95%CI,68.8%-88.2%）,and the specificity was 83.42%（95%CI,82.3%-84.5%）.（5）Comparison of the optimal HbA1c cut-off point and its efficiency based on different reference standardsThe diagnostic cut-off point based on specific DR（6.2%）was higher than that based on the WHO standard（6.1%）,but there was no significant difference in the AUC between the different standards（P=0.1727）.Conclusion:1.The optimal diagnostic cut-off point for HbA1c was at 6.0%and around 6.3%based on the OGTT results and specific DR,respectively.2.An increased risk of misdiagnosis may exist when using the same HbA1c threshold in the elderly,overweight and obese,high LDL-C,the Dai,or the Uyghur population.3.Based on the WHO standard,the optimal HbA1c cut-off points of each ethnic group were close to 6.0%,except for the Dai（6.3%）.Besides,the older the age,the higher the HbA1c optimal cut-off point.Based on the specific DR,HbA1c optimal cut-off points were different varying with ethnicity,gender,age,and BMI.4.The efficiency of HbA1c combined with fasting blood glucose in diagnosing diabetes was higher than or equal to the OGTT.Which provided evidence that HbA1c combined with fasting blood glucose can be used as a substitute for OGTT in the future.Part 2 Evaluation of the artificial intelligence-based pattern recognition system for diabetic retinopathy screening in China Background and aims: Diabetic retinopathy（DR）is the leading cause of vision loss among working-age adults.The International Diabetes Federation recommends that the moderate and worse diabetic retinopathy patients require timely referral and treatment.Regular screening and early detection are keys to manage DR.Recently,artificial intelligence（AI）has gained wide acceptance as a virtual assistant in DR detection,and promoted the health care more efficient,accessible and cost-effective.Studies from the USA,Singapore,and China have reported that the sensitivity of AI system is unstable ranging from 79%-98% in the DR screening in the diabetes population.It is the first time to evaluate the diagnostic accuracy of the AI-based pattern recognition system in detecting diabetic retinopathy in the general population.Materials and methods: The cross-sectional database was used to analyze from the cohort study conducted among a randomized cluster sample of 18-70 years old residents in different ethnic groups in China.For each enrolled participant,2 retinal fundi（one disc and one macula centered 45°）images per eye were collected for the evaluation of DR.All of the images were graded by two ophthalmologists using the Early Treatment Diabetic Retinopathy Study（ETDRS）scale,and images of disc were submitted to the AI system for grading.ETDRS was used to diagnose DR as the gold standard.Referable DR was defined as moderate and worse diabetic retinopathy（ETDRS>31）.Any DR was defined as ETDRS≥20.Participant defined as DR was with images of DR in any fundus.Area under the receiver operating characteristic curve（AUC）,sensitivity and specificity of the AI-based pattern recognition were performed by the receiver operating characteristic curve（ROC）based on the ETDRS using the results from every individual person.Results: 17412 images from 9037 participants（mean [SD] age,50.1[12.0] years old）were graded.In the general population,for referable DR,the AUC,sensitivity and specificity was0.941（95% CI,0.936 to 0.946）,98.15%（95% CI,90.1%-100.0%）,and 90.06%（95% CI,89.4%-90.7%）,respectively.For any DR,AUC was 0.881（95% CI,0.874 to 0.887）,sensitivity was 83.62%（95% CI,79.3%-87.4%）,and specificity was 92.49 %（95%CI,91.9%-93.0%）.In the diabetes group,for referable DR,the AUC,sensitivity and specificity was 0.901（95% CI,0.883 to 0.917）,97.83%（95% CI,88.5%-99.9%）,and 82.43%（95% CI,80.1%-84.6%）,respectively.For any DR,the AUC,sensitivity and specificity was 0.900（95% CI,0.882 to 0.917）,91.10%（95% CI,85.3%-95.2%）,88.96%（95% CI,86.9%-90.8%）,respectively.Conclusion: The AI-based pattern recognition system showed high sensitivity and specificity in detecting referable DR both in the general population and diabetic group,based on the ETDRS results from the ophthalmologists.Further study is necessary to evaluate whether the AI system can improve DR outcomes compared with the clinical doctors,and find the relative factors disturbing the AI detection.