Original article
German Diabetes Risk Score for the Determination of the Individual Type 2 Diabetes Risk
10-year prediction and external validations
; ; ; ; ; ;
Background: The German Diabetes Risk Score (GDRS) currently enables prediction of the individual risk of developing type 2 diabetes (T2D) within five years. The aim of this study is to extend the prediction period of the GDRS, including its non-clinical version and its HbA1c extension, to 10 years, and to perform external validation.
Methods: In data from the European Prospective Investigation into Cancer and Nutrition (EPIC)-Potsdam study (n = 25 393), Cox proportional hazards regression was used to reweight the points that were used to calculate the five-year risk. Two population-based prospective cohorts (EPIC-Heidelberg n = 23 624, GNHIES98 cohort n = 3717) were used for external validation. Discrimination was represented by C-indices, and calibration by calibration plots and the expected-to-observed (E/O) ratio.
Results: Prediction performance in EPIC-Potsdam was very good (C-index for the non-clinical model: 0.834) and was confirmed in EPIC-Heidelberg (0.843) and in the GNHIES98 cohort (0.851). Among persons in the GNHIES98 cohort with a greater than 10% predicted probability of disease, 14.9% developed T2D within 10 years (positive predictive value). The models were very well calibrated in EPIC-Potsdam (E/O ratio for the non-clinical model: 1.08), slightly overestimated the risk in EPIC-Heidelberg (1.34), and predicted T2D very well in the GNHIES98 cohort after recalibration (1.06).
Conclusion: The extended GDRS prediction period of 10 years, with a non-clinical version and an HbA1c extension that will soon be available in both German and English, enables the even longer-range, evidence-based identification of high-risk individuals with many different applications, including medical screening.
Risk-scoring systems for type 2 diabetes (T2D) calculate a statistical estimate of individual disease risk, based on selected parameters, the so-called predictors. Risk scores can be used to identify persons who are at high risk and may benefit particularly from prevention measures. The German Diabetes Risk Score (GDRS) developed by the German Institute of Human Nutrition Potsdam-Rehbrücke quantifies the individual 5-year risk of T2D (1, 2). The score is based on anthropometric data, and information on lifestyle and family history of T2D and can thus be used for objective individual risk estimation both in non-clinical preventive settings and in doctor–patient conversations. A simplified paper questionnaire has been developed for calculating the risk in the absence of digital devices, and if the HbA1c concentration is known there is also a HbA1c extension that further improves the accuracy of risk prediction (3, 4).
The prediction period was previously limited to 5 years, in line with the duration of follow-up in the underlying European Prospective Investigation into Cancer and Nutrition (EPIC)-Potsdam study. In the meantime, however, the longer follow-up period has enabled revision of the GDRS to predict the 10-year risk.
The aim of the work presented here was to extend the GDRS to predict the individual 10-year risk and to externally validate the performance in two independent population-based studies.
Methods
Study populations
The analyses were based on data from the two German EPIC cohorts [EPIC-Potsdam and EPIC-Heidelberg] together with longitudinal data from the German National Health Interview and Examination Survey 1998 (GNHIES98 cohort).
The EPIC-Potsdam and EPIC-Heidelberg cohorts contained 27 548 and 25 540 participants, respectively, from the areas around Potsdam (mean age 50 years, standard deviation [SD] 9 years; 60.4% female) and Heidelberg (mean age 50 years, SD 8 years; 53.3% female). The data acquisition procedures have already been described (5, 6). Briefly, baseline assessment (1994–1998) of the participants comprised physical examination, questioning on lifestyle, sociodemographic characteristics, and disease status using validated questionnaires and in personal interviews, together with analysis of blood samples taken by trained study staff (eMethods 1). In both cohorts, participants with self-reported earlier diagnosis of T2D by a physician, non-verifiable T2D in the follow-up period, or other forms of diabetes were excluded, as were those for whom follow-up data were completely missing. The analyzed samples were composed of 25 393 persons in Potsdam and 23 624 persons in Heidelberg (eMethods 2). The studies were approved by the ethics committees of the federal state of Brandenburg and the University of Heidelberg.
The GNHIES98 cohort comprised 3959 persons who had taken part in the nationwide interview and examination surveys GNHIES98 (1997–1999) and in the German Health Interview and Examination Survey for Adults (DEGS)1 (2008–2011) (age 18–79 years, 51.6% female). The design has been described previously (7). In brief, GNHIES98 and DEGS1 featured data acquisition by means of standardized physical examination and blood sampling by trained study staff, a self-completed questionnaire, standardized computer-aided interviews by physicians, and drug interviews (eMethods 1). Participants with diagnosed diabetes or missing data on diabetes at the time of baseline examination; missing diabetes data, type 1 diabetes, or gestational diabetes at follow-up; or missing or implausible information on age at the time of diabetes diagnosis during follow-up were excluded (eMethods 2). The analyzed sample contained 3717 persons. DEGS1 was approved by the ethics committee of Charité University Medical Center Berlin (no. EA2/047/08).
For the development and validation of the HbA1c extension, we excluded from all three cohorts already diseased but undiagnosed participants with a HbA1c level ≥ 6.5% at baseline examination (eMethods 2).
Participants were told about the studies’ goals and procedures as well as data protection and gave their written informed consent.
Identification and definition of T2D cases
In EPIC-Potsdam and EPIC-Heidelberg, potential incident cases of T2D—i.e., participants who fell ill after the baseline examination—were systematically documented with regard to self-reporting of a diagnosis, diabetes-relevant medications, and diabetes-related advice to modify diet during the follow-up period (EPIC-Potsdam: up to January 2011; EPIC-Heidelberg: up to August 2015). Furthermore, incidental diagnoses of T2D from other sources such as hospitals, physicians, death certificates, and tumor centers were recorded (6). Potential cases of T2D were systematically checked by means of verification forms sent to the treating physicians. Only confirmed new cases of T2D (code E11 in the International Classification of Diseases, 10th revision) were included for analysis.
In the GNHIES98 cohort, cases identified after baseline examination were systematically identified up to December 2011. To that end, self-reports of diagnosis by a physician were identified in standardized computer-aided interviews with a physician or by documentation of the intake of diabetes-related medications in the 7-day period preceding the interview through automated drug documentation and with an indicated diagnosis year after baseline examination (8) (eMethods 3).
Statistical methods
In EPIC-Potsdam, EPIC-Heidelberg, and the GNHIES98 cohort, missing data in the predictors of the baseline examination were estimated by multiple imputation (eMethods 4; comparison of non-imputed and imputed values in [9]).
Development of the model in EPIC-Potsdam was based on the whole follow-up period. Observations were censored at the time of T2D diagnosis, the last available follow-up information, or death. Two options for model adjustment were pursued by means of Cox regression: On the one hand, the previous GDRS score points were modeled as a variable (estimation of a weight and the survival function, which describes the likelihood that a T2D diagnosis has not occurred within 10 years); on the other hand, the individual GDRS predictors were modeled (re-estimation of the individual weights and estimation of the survival function after 10 years) (eMethods 5). The censored time until T2D diagnosis served as dependent variable. In the following step, the model performance for a follow-up period of 10 years was determined in EPIC-Potsdam and compared. To create a paper questionnaire for risk estimation independent of electronic devices, the prediction parameters were categorized and scoring points were assigned (4). The sum of the GDRS questionnaire points was modeled as a variable—including estimation of a weight and the survival function. The HbA1c extension was newly estimated for a 10-year prediction period, including weighting of the GDRS score points and HbA1c.
For purposes of external validation the model performance for a 10-year prediction period was determined in EPIC-Heidelberg (HbA1c extension only in the subcohort) and in the GNHIES98 cohort.
The performance was quantified in terms of discrimination and calibration. For discrimination, Harrell’s C-index was quantified in EPIC-Potsdam and EPIC-Heidelberg and the area under the receiver operating characteristic curve (ROC-AUC) in the GNHIES98 cohort (eMethods 5), in each case together with sensitivity, specificity, negative predictive value (NPV), and positive predictive value (PPV) (thresholds: 5%, 7.5%, 10% [10]). For calibration, we generated calibration plots based on deciles of the predicted risk and calculated the expected-to-observed ratio (E/O ratio). Because the calibration is determined largely by sample-specific incidence, the prediction models were recalibrated in the GNHIES98 cohort, as previously described (3).
All statistical analyses were performed using SAS (version 9.4).
Results
The mean age of the samples in the two EPIC cohorts was 50 years (SD: EPIC-Potsdam 9 years, EPIC-Heidelberg 8 years) and in the GNHIES98 cohort 43 years (SD: 25 years) (Table 1). Female participants predominated in EPIC-Potsdam (61.4%), while the sexes were more equally represented in EPIC-Heidelberg (54.3% female) and the GNHIES98 cohort (50.9% female). In EPIC-Potsdam 1563 cases of T2D were detected (mean duration of follow-up 10.5 years, SD 2.5 years), 1367 of them in the first 10 years; in EPIC-Heidelberg, 1223 cases in the first 10 years (mean follow-up 9.2 years, SD 2.1 years); in the GNHIES98 cohort, 179 cases in the first 10 years (mean follow-up 9.8 years, SD 1.2 years) (Table 1).
Revised models and performance in EPIC-Potsdam
A performance comparison between the two adjusted non-clinical GDRS models is presented in eResults 1. Modeling of the GDRS predictors individually yielded no advantage over modeling of the previous GDRS score points, so the latter seemed adequate for updating of the GDRS.
The equations used to calculate the absolute risk with the non-clinical GDRS, the simplified version of the questionnaire, and the HbA1c extension are contained in the supplementary material (Figure 2, eResults 2, sample calculation eResults 3).
The non-clinical GDRS discriminated very well in EPIC-Potsdam (C-index 0.834). The simplified questionnaire discriminated comparably well, while the HbA1c extension attained an even higher C-index (0.853) (Figure 1). The C-indices were higher in women than in men and higher in the second half of the follow-up period (> 5 years) than in the first (eResults 4). For a threshold of 7.5%, for example, the weighted GDRS model showed a sensitivity of 68.7%, specificity of 80.2%, PPV of 16.5% (i.e., 16.5% of persons with a predicted risk of > 7.5% were diagnosed with T2D in the following 10 years), and NPV of 97.8% (eResults 5).
With regard to calibration, the non-clinical GDRS (E/O ratio with 95% confidence intervals [CI]: 1.08 [1.03; 1.14]), the simplified version of the questionnaire (1.09 [1.04; 1.15]), and the HbA1c extension (1.10 [1.04; 1.17]) all performed very well (eResults 6, Table 2).
External validation in EPIC-Heidelberg and the GNHIES98 cohort
In EPIC-Heidelberg the C-indices were slightly higher than in EPIC-Potsdam and confirmed the very good discrimination (95% CI for C-index: non-clinical GDRS 0.843 [0.811; 0.872], simplified GDRS 0.846 [0.815; 0.875], HbA1c extension 0.862 [0.757; 0.943]; Figure 1). For a threshold of 7.5%, for example, the GDRS in EPIC-Heidelberg showed a sensitivity of 76.2%, specificity of 75.9%, PPV of 14.7% and NPV of 98.3% (eResults 5).
In the GNHIES98 cohort the C-indices were higher than in the EPIC cohorts (non-clinical GDRS 0.851 [0.826; 0.877], simplified GDRS 0.849 [0.822; 0.876], HbA1c extension 0.883 [0.854; 0.912]; Figure 1). When only including persons comparable in age with the EPIC-Potsdam participants (35–65 years), the discrimination was somewhat weaker but still very good (non-clinical GDRS 0.822 [0.790; 0.852], simplified GDRS 0.816 [0.782; 0.848], HbA1c extension 0.878 [0.843; 0.909]) (for sensitivity, specificity, PPV, and NPV, see eResults 5).
With regard to calibration, in EPIC-Heidelberg the risk was overestimated by the non-clinical GDRS and the simplified GDRS (E/O ratio: non-clinical GDRS 1.34, simplified GDRS 1.35; Table 2). In the lower nine deciles of predicted risk, there was only a slight overestimation which was however more pronounced in the highest decile. (eResults 6). The HbA1c extension was very well calibrated with only minor deviations (0.99).
The GDRS models also overestimated the risk in the GNHIES98 cohort (eResults 7–9). After recalibration the performance was good to very good for the non-clinical GDRS (E/O ratio 1.06), the simplified GDRS (1.12), and the HbA1c extension (1.13) (eResults 6, Table 2).
Discussion
This study has extended the prediction period of the nonclinical versions of the GDRS and the HbA1c extension from 5 years to 10 years. The very good discrimination achieved in EPIC-Potsdam was confirmed by two external validations in EPIC-Heidelberg and the GNHIES98 cohort. Although the models were very well calibrated in EPIC-Potsdam, they overestimated the risk at the higher risk range in EPIC-Heidelberg and the GNHIES98 cohort. After recalibration, prediction in the GNHIES98 cohort was very accurate.
In previous studies, the discrimination achieved by the 5-year GDRS in EPIC-Potsdam, EPIC-Heidelberg, and the GNHIES98 cohort, despite the shorter prediction period, lay only just above, or even below, the discrimination of the models we developed for 10-year prediction (e.g., EPIC-Potsdam C-index, 95% CI: 0.83 [0.81; 0.84]), the simplified version of the questionnaire (0.83), and the HbA1c extension (0.87 [0.81; 0.92]) (1, 2, 3, 4). In agreement with the findings of our study, owing to the wider age range in the GNHIES98 cohort, the earlier studies on 5-year risk also found somewhat higher discrimination in the GNHIES98 cohort (18–79 years) than in EPIC-Potsdam (35–65 years) (11). Similarly, assessment of calibration in previous studies on the 5-year GDRS also showed slight overestimation of risk in the GNHIES98 cohort and the MONICA/KORA study (2, 3, 11). This may be attributable to genuine differences in incidence rates or to different survey procedures for identification of T2D cases.
Various aspects of this study are worthy of note. As for the 5-year GDRS, three versions of the model were developed: the non-clinical GDRS, the simplified version of the questionnaire, and the HbA1c extension (see decision tree in Figure 3). Together, these cover a wide spectrum for prediction of the 10-year risk. They can be used clinically in routine screening examinations; for self-information; when clinical parameters are not available; and at public health level in public awareness campaigns and potentially for population-wide screening programs. As with the 5-year GDRS, screening could be a multistage process. In a first step, the non-clinical versions could be used for initial rapid and resource-sparse risk assessment based on non-invasive and readily determined parameters. To this end, if digital devices are available for risk calculation we recommend the full non-clinical GDRS model, as despite the comparable very good prediction quality it cannot be excluded that the categorization of the continuous prediction parameters for the simplified version of the questionnaire will lead to information loss. In a second step, HbA1c measurement can be used to achieve even more accurate prediction of the risk.
The German Diabetes Society (Deutsche Diabetes Gesellschaft) recommends risk scores, the 5-year GDRS among others, for primary T2D screening as indicators for blood sugar determination in the context of diagnosis (12). Nevertheless, to the best of our knowledge there exists no overarching prevention strategy that has been promulgated by professional societies and which, guided by the predicted individual risk, provides for systematic, evidence-based, timely preventive measures such as lifestyle interventions to reduce the risk.
As far as we know, the models presented here are the first for prediction of the 10-year risk of T2D to be explicitly developed on the basis of a German sample. The fact that the predictors overlap with an externally validated test for estimation of the 10-year risk of cardiovascular disease (CVD) facilitates simultaneous calculation of the risks of T2D and CVD (9). The existing GDRS online tool (non-clinical GDRS) (13) and the paper questionnaires (simplified questionnaire and HbA1c extension) have already been modified for prediction of the 10-year risk and are available in German and in English for straightforward and user-friendly application.
Our study has a number of strengths. The prediction models were developed in a population-based prospective study, the preferred design for model development (14). Moreover, the models underwent two external validation processes: in a comprehensive population-based observational study (EPIC-Heidelberg) and in a sample representative for Germany (the GNHIES98 cohort). These studies have relevant sample sizes, and the extensive verification procedures of diabetes cases in the EPIC samples minimizes the rate of false-positive diagnoses.
However, the study also has limitations:
- The survey procedures differed slightly between the EPIC cohorts and the GNHIES98 cohort, e.g, in the scaling of the data on exercise and the identification of new cases of diabetes. Nevertheless, the prediction quality in the GNHIES98 cohort was very good, indicating relative robustness of the models to slight deviations in data acquisition.
- Owing to the case–cohort design, the proportion of missing data on biomarker measurement was relatively high in the EPIC populations. However, an earlier study showed that multiple imputation is a valid method for estimation of missing data in the risk prediction context (15).
- The family history of diabetes was documented only for the participants’ parents in the GNHIES98 cohort and only at the fifth follow-up visit in the EPIC samples, which may have resulted in slight overestimation of the rate of family members with T2D at the time of baseline examination.
- The reparticipation rate of GNHIES98 participants in the DEGS survey was moderate, at 62%. Despite an analytical weighting factor (16), selection bias cannot be excluded (8).
- Due to the cut-offs used, the PPV was relatively low. However, an earlier study on the 5-year GDRS showed that lifestyle interventions are cost-benefit-efficient even at low threshold values from the point of view of health economics (17). Equivalent cost–benefit analyses on the 10-year GDRS are needed for determination of ideal intervention thresholds.
- The aim of this study was to extend the GDRS prediction period. Other, potentially informative parameters of prediction were not investigated.
- In common with other T2D scores, the models estimate the risk of a T2D diagnosis, not the preceding onset of the disease (18).
Conclusions
The versions of the GDRS developed to predict the 10-year risk of T2D (Figure 3) enable evidence-based, accurate estimation of the individual risk in various circumstances. They can be used in persons without type 2 or type 1 diabetes in the age range 18 to 79 years, e.g., in a screening examination. This enables identification of persons at elevated risk who might benefit from preventive lifestyle interventions to avoid or delay the onset of T2D.
Funding
This work was funded by the German Federal Ministry of Science [01 EA 9401], the European Union [SOC 95201408 05F02, SOC 98200769 05F02], German Cancer Aid [70–2488-Ha I], the German Federal Ministry of Education and Research [01ER0808; 01ER0809], and the federal state of Brandenburg via the German Center for Diabetes Research [82DZD00302] and the German Center for Cancer Research (DKFZ). The studies GNHIES98 and DEGS1 were financed and supported by the Federal Ministry of Health and the Robert Koch Institute [GE20190305].
Conflict of interest statement
Prof. Fritsche has received lecture fees from Sanofi, Novo Nordisk, Astra Zeneca, and Boehringer Ingelheim.
The remaining authors declare that no conflict of interest exists.
Manuscript received on 14 February 2022, revised version accepted on 30 June 2022.
Translated from the original German by David Roseveare
Corresponding author
Dr. rer. nat. Catarina Schiborn
Abteilung Molekulare Epidemiologie
Deutsches Institut für Ernährungsforschung Potsdam-Rehbrücke (DIfE)
Arthur-Scheunert-Allee 114–116
14558 Nuthetal, Germany
catarina.schiborn@dife.de
Cite this as:
Schiborn C, Paprott R, Heidemann C, Kühn T, Fritsche A, Kaaks R, Schulze MB: German diabetes risk score for the determination of the individual type 2 diabetes risk—10-year prediction and external validations. Dtsch Arztebl Int 2022; 119: 651–7. DOI: 10.3238/arztebl.m2022.0268
►Supplementary material:
eMethods, eResults
www.aerzteblatt-international.de/m2022.0268
German Center for Diabetes Research (DZD), Munich: Dr. rer. nat. Catarina Schiborn, Prof. Dr. med. Andreas Fritsche, Prof. DrPH Matthias B. Schulze
Department of Epidemiology and Health Monitoring, Robert Koch Institute (RKI), Berlin: Dr. oec. troph. Rebecca Paprott, DrPH Christin Heidemann
Division of Cancer Epidemiology, German Cancer Research Center (DKFZ), Heidelberg: PD Dr. sc. hum. Tilman Kühn, Prof. PhD Rudolf Kaaks
Institute for Global Food Security, Queen’s University Belfast, Belfast, UK: PD Dr. sc. hum. Tilman Kühn
Department of Medicine IV, University Hospital Tübingen: Prof. Dr. med. Andreas Fritsche
Institute for Diabetes Research and Metabolic Diseases of the Helmholtz Center Munich at the University of Tübingen, Tübingen: Prof. Dr. med. Andreas Fritsche
Institute of Nutritional Science, University of Potsdam, Nuthetal: Prof. DrPH Matthias B. Schulze
| 1. | Schulze MB, Hoffmann K, Boeing H, et al.: An accurate risk score based on anthropometric, dietary, and lifestyle factors to predict the development of type 2 diabetes. Diabetes Care 2007; 30: 510–5 CrossRef CrossRef CrossRef MEDLINE |
| 2. | Muhlenbruch K, Ludwig T, Jeppesen C, et al.: Update of the German Diabetes Risk Score and external validation in the German MONICA/KORA study. Diabetes Res Clin Pract 2014; 104: 459–66 CrossRef MEDLINE |
| 3. | Muhlenbruch K, Paprott R, Joost HG, Boeing H, Heidemann C, Schulze MB: Derivation and external validation of a clinical version of the German Diabetes Risk Score (GDRS) including measures of HbA1c. BMJ Open Diabetes Res Care 2018; 6: e000524 CrossRef MEDLINE PubMed Central |
| 4. | Schulze MB, Holmberg C, Hoffmann K, Boeing H, Joost HG: Kurzfragebogen zur Bestimmung des Diabetesrisikos auf Grundlage des Deutschen Diabetes-Risiko-Scores. Ernährungs Umschau 2007; 54: 698–703. |
| 5. | Boeing H, Korfmann A, Bergmann MM: Recruitment procedures of EPIC-Germany. European Investigation into Cancer and Nutrition. Ann Nutr Metab 1999; 43: 205–15 CrossRef MEDLINE |
| 6. | Bergmann MM, Bussas U, Boeing H: Follow-up procedures in EPIC-Germany—data quality aspects. European Prospective Investigation into Cancer and Nutrition. Ann Nutr Metab 1999; 43: 225–34 CrossRef MEDLINE |
| 7. | Scheidt-Nave C, Kamtsiuris P, Gößwald A, et al.: German health interview and examination survey for adults (DEGS)—design, objectives and implementation of the first data collection wave. BMC Public Health 2012; 12: 730 CrossRef MEDLINE PubMed Central |
| 8. | Heidemann C, Niemann H, Paprott R, Du Y, Rathmann W, Scheidt-Nave C: Residential traffic and incidence of Type 2 diabetes: the German Health Interview and Examination Surveys. Diabet Med 2014; 31: 1269–76 CrossRef MEDLINE |
| 9. | Schiborn C, Kühn T, Mühlenbruch K, et al.: A newly developed and externally validated non-clinical score accurately predicts 10-year cardiovascular disease risk in the general adult population. Sci Rep 2021; 11: 19609 CrossRef MEDLINE PubMed Central |
| 10. | Goff DC, Lloyd-Jones DM, Bennett G, et al.: 2013 ACC/AHA guideline on the assessment of cardiovascular risk: a report of the American College of Cardiology/American Heart Association Task Force on Practice Guidelines. J Am Coll Cardiol 2014; 63: 2935–59 CrossRef MEDLINE PubMed Central |
| 11. | Paprott R, Muhlenbruch K, Mensink GB, et al.: Validation of the German Diabetes Risk Score among the general adult population: findings from the German Health Interview and Examination Surveys. BMJ Open Diabetes Res Care 2016; 4: e000280 CrossRef MEDLINE PubMed Central |
| 12. | Schleicher E, Gerdes C, Petersmann A, et al.: Definition, Klassifikation und Diagnostik des Diabetes mellitus: Update 2021. Diabetologie und Stoffwechsel 2021; 16: 110–8 CrossRef |
| 13. | Deutsches Institut für Ernährungsforschung Potsdam-Rehbrücke: DIfE – DEUTSCHER DIABETES-RISIKO-TEST. drs.dife.de (last accessed on 4 August 2022). |
| 14. | Moons KG, Kengne AP, Woodward M, et al.: Risk prediction models: I. Development, internal validation, and assessing the incremental value of a new (bio)marker. Heart 2012; 98: 683–90 CrossRef MEDLINE |
| 15. | Muhlenbruch K, Kuxhaus O, di Giuseppe R, Boeing H, Weikert C, Schulze MB: Multiple imputation was a valid approach to estimate absolute risk from a prediction model based on case-cohort data. J Clin Epidemiol 2017; 84: 130–41 CrossRef MEDLINE |
| 16. | Kamtsiuris P, Lange M, Hoffmann R, et al.: Die erste Welle der Studie zur Gesundheit Erwachsener in Deutschland (DEGS1). Bundesgesundheitsblatt – Gesundheitsforschung – Gesundheitsschutz 2013; 56: 620–30 CrossRef MEDLINE |
| 17. | Muhlenbruch K, Zhuo X, Bardenheier B, et al.: Selecting the optimal risk threshold of diabetes risk scores to identify high-risk individuals for diabetes prevention: a cost-effectiveness analysis. Acta Diabetol 2020; 57: 447–54 CrossRef MEDLINE PubMed Central |
| 18. | Porta M, Curletto G, Cipullo D, et al.: Estimating the delay between onset and diagnosis of type 2 diabetes from the time course of retinopathy prevalence. Diabetes Care 2014; 37: 1668–74 CrossRef MEDLINE |
