Diabetes Care 1

Impact of Level on Micro- Frida Emanuelsson,1,2,3 Sarah Marott,1,2,3 Anne Tybjærg-Hansen,1,2,3,4 and Macrovascular Disease in the Børge G. Nordestgaard,2,3,4,5 and Marianne Benn1,2,3 General Population: A Mendelian Randomization Study https://doi.org/10.2337/dc19-1850

OBJECTIVE To evaluate whether high glucose levels in the normoglycemic range and higher have a causal genetic effect on risk of retinopathy, neuropathy, nephropathy, chronic kidney disease (CKD), peripheral arterial disease (PAD), and myocardial infarction (MI; positive control) in the general population.

RESEARCH DESIGN AND METHODS This study applied observational and one-sample Mendelian randomization (MR) analyses to individual-level data from 117,193 Danish individuals, and validation by two-sample MR analyses on summary-level data from 133,010 individuals from the Meta-Analyses of Glucose and -Related Traits Consortium (MAGIC), 117,165 1Department of Clinical Biochemistry, Rigshospi- ADOACLRADMTBLCRISK METABOLIC AND CARDIOVASCULAR from the CKDGen Consortium, and 452,264 from the UK Biobank. talet, Copenhagen University Hospital, Copenha- gen, Denmark 2 RESULTS Faculty of Health and Medical Sciences, Univer- sity of Copenhagen, Copenhagen, Denmark Observationally, glucose levels in the normoglycemic range and higher were as- 3TheCopenhagenGeneralPopulationStudy,Herlev sociatedwith high risksof retinopathy, neuropathy,, PAD,and and Gentofte Hospital, Copenhagen University P < Hospital, Herlev, Denmark MI (all for trend 0.001). In genetic causal analyses, the risk ratio for a 1 mmol/L 4 – – The Copenhagen City Heart Study, Frederiksberg higher glucose level was 2.01 (95% CI 1.18 3.41) for retinopathy, 2.15 (1.38 3.35) for and Bispebjerg Hospital, Copenhagen University neuropathy, 1.58 (1.04–2.40) for diabetic nephropathy, 0.97 (0.84–1.12) for esti- Hospital, Frederiksberg, Denmark mated glomerular filtration rate (eGFR) <60 mL/min/1.73 m2, 1.19 (0.90–1.58) for 5Department of Clinical Biochemistry, Herlev PAD, and 1.49 (1.02–2.17) for MI. Summary-level data from the MAGIC, the CKDGen and Gentofte Hospital, Copenhagen University Hospital, Herlev, Denmark Consortium, and the UK Biobank gave a genetic risk ratio of 4.55 (95% CI 2.26–9.15) – – Corresponding author: Marianne Benn, marianne for retinopathy, 1.48 (0.83 2.66) for peripheral neuropathy, 0.98 (0.94 1.01) for [email protected] eGFR <60 mL/min/1.73 m2, and 1.23 (0.57–2.67) for PAD per 1 mmol/L higher Received 17 September 2019 and accepted 23 glucose level. January 2020 This article contains Supplementary Data online CONCLUSIONS at http://care.diabetesjournals.org/lookup/suppl/ Glucose levels in the normoglycemic range and higher were prospectively asso- doi:10.2337/dc19-1850/-/DC1. ciated with a high risk of retinopathy, neuropathy, diabetic nephropathy, eGFR © 2020 by the American Association. <60 mL/min/1.73 m2, PAD, and MI. These associations were confirmed in ge- Readers may use this article as long as the work is properly cited, the use is educational and not for netic causal analyses for retinopathy, neuropathy, diabetic nephropathy, and MI, fi fi < pro t, and the work is not altered. More infor- but they could not be con rmed for PAD and seemed to be refuted for eGFR 60 mation is available at https://www.diabetesjournals 2 mL/min/1.73 m . .org/content/license. Diabetes Care Publish Ahead of Print, published online February 13, 2020 2 Impact of Glucose Level on Vascular Disease Diabetes Care

High glucose levels are observationally glucose levels (1,9,10); thus, second, we were used in order to genotype seven and causally associated with a high risk of tested whether they also did so in our variants of the genes GCP62/ABCB1 ischemic heart disease in individuals with populations. Third, we tested whether (rs560887), GCK (rs4607517), DGKB and those without diabetes (1,2). Simi- the genetic variants associated with high (rs2191349), ADCY5 (rs11708067), CDKN2A/ larly, is a well-known risk glucose levels also associated, as an in- B (rs10811661 and rs2383206), and TCF7L2 factor for microvascular disease in indi- dication of causality, with risk of disease. (rs7903146), which have been shown to viduals with diabetes, and treatment with Fourth, we analyzed instrumental varia- be associated with high glucose levels glucose-lowering drugs reduces the risk bles to obtain causal risk estimates per (1,9,10). The variants were combined into of microvascular complications in such 1 mmol/L (18 mg/dL) higher glucose a weighted allele score, which accounted individuals (3,4). Whether high glucose levels. Fifth and last, we validated our for the effect of each allele on random levels increase the risk of a spectrum of results using a two-sample MR design plasma glucose and the frequency of peripheral micro- and macrovascular dis- with summary-level data based on 26 ge- each allele in the CCHS and CGPS pop- eases in the general population, and netic variants associated with fasting glu- ulations (16). In order to compare the whether the magnitude of such risk varies cose levels in individuals without diabetes effect of the same genetic variants on between vascular compartments, is less from the Meta-Analyses of Glucose and fastingglucoselevels, theallele scorewas clear. Also uncertain is whether the ef- Insulin-RelatedTraitsConsortium(MAGIC), also weighted on the basis of its effect on fect of glucose-lowering drugs on vascu- and end point data from the UK Biobank fasting glucose levels in MAGIC partic- lar disease is caused by glucose lowering and the CKDGen Consortium (11–14). ipants. The weighted allele scores were per se or whether other mechanisms also divided into five categories, cut at the are involved, as most glucose-lowering RESEARCH DESIGN AND METHODS 25th, 50th, 75th, and 95th percentiles, in fl drugs have pleiotropic effects, including Study Population order to re ect stepwise increases in glu- improvements in concomitant obesity, We included 117,193 individuals from cose levels from below the median to hyperlipidemia, and hypertension (5,6). two similar studies of the Danish general extremely high values. In order to determine whether a risk population: the CCHS and the CGPS (1). factor is causally related to a disease, Participants were white and of Danish End Points several lines of evidence should be con- descent, and none were included in more End points were based on diagnoses of sidered, including mechanistic studies, lon- than one study. For detailed descrip- retinopathy, peripheral neuropathy, di- gitudinal epidemiological studies, genetic tions of the cohorts, see Supplementary abetic nephropathy, PAD, and MI accord- Mendelian randomization (MR) studies, Appendix 1. Both studies were approved ing to the codes from the World Health and randomized controlled trials (7,8). by institutional boards at Herlev and Gen- Organization’s ICD-8 and ICD-10. (For spe- MR is an epidemiological approach to tofte Hospitals and The Danish National cific ICD codes, see Supplementary Table assessing causality that uses the random Committee on Health Research Ethics, 1.) Relevant data were collected from assortment of alleles at conception to Copenhagen, Denmark (KF-100.2039/91, 1 January 1977 through 10 April 2018 largely circumvent confounding and re- KF-01–144/01, H-KF-01–144/01), and they by reviewing diagnoses from all hospital verse causation (8). Applying the princi- were conducted according to the Declara- admissions and outpatient clinic visits in ples of MR in a general population, we tion of Helsinki. Written informed consent the national Danish Patient Registry and tested the hypothesis that high glucose was obtained from all individuals. the national Danish Registry of Causes of levels within the normal range and higher Death. CKD was defined as an estimated causally associate with high risks of ret- Plasma Glucose glomerular filtration rate (eGFR) ,60 inopathy, peripheral neuropathy, and Nonfasting plasma glucose was measured mL/min/1.73 m2, corresponding to the diabetic nephropathy, representing mi- by using a colorimetric assay (Konelab; KDIGO definition of mildly to moderately crovascular disease, or with high risks of Boehringer, Mannheim, Germany). Blood decreased (stage G3) kidney function or chronic kidney disease (CKD) and periph- samples were taken at random, irrespective worse (17). eGFR was calculated from eral arterial disease (PAD), representing a of time since and content of the last meal. creatinine by using the CKD-EPI creati- mix of micro- and macrovascular PAD. Time (hours) since the last meal was self- nine equation (18). For detailed descrip- Myocardial infarction (MI) was included reported. In order to examine the observa- tions of end points and covariates, see as a positive control for the genetic in- tional association between glucose levels Supplementary Appendix 2. strument, that is, as a confirmation of the and disease, individuals without diabetes association between the genetic instru- were categorized, on the basis of glucose Statistical Analyses ment and the risk of MI. level at baseline, into five groups that reflect We used Stata SE 14.2 to analyze the data. In two cohorts from the Danish general stepwise increases in glucose level, ranging Deviation from Hardy-Weinberg equilib- populationdparticipants in the Copen- from $4.0 mmol/L to above the diabetes rium was tested by using the Pearson hagen City Heart Study (CCHS) and those threshold of 11.1 mmol/L (15). Individuals x2 test. To test for trend across ordered in the Copenhagen General Population with levels between 4.0 and 5.4 mmol/L categories of glucose levels, genotypes, Study (CGPS)dwe first tested whether were defined as the reference group. and weighted allele scores, we used random plasma glucose levels at baseline the nonparametric Cuzick extension of a were prospectively associated with risk of Genotypes Wilcoxon rank sum test. disease. Seven variants of the genes GCP62/ An ABI PRISM 7900HT Sequence Detec- First, the observational associations be- ABCB11, GCK, DGKB, ADCY5, CDKN2A/B, tion System (Applied Biosystems Inc., Foster tween glucose level and disease in indi- and TCF7L2 have been shown to affect City, CA) and TaqMan-based assays viduals without diabetes were examined care.diabetesjournals.org Emanuelsson and Associates 3

for predefined categories of glucose level the genetic instrument should influence at baseline or a glucose level below 4.0 (which increases with each category) by the risk of disease only through the ex- mmol/L, leaving 112,455 individuals for using Cox proportional hazards regres- posure of interest (i.e., plasma glucose). To the observational analyses. Baseline char- sion with age as the time scale and with test this, we used logistic regression to acteristics for all individuals, and for in- delayedentry (left-truncation), excluding assess whether the potential confoundersd dividuals without diabetes by glucose individuals with a glucose level ,4.0 age, sex, current smoking, BMI, LDL cho- category, are shown in Table 1. Mean mmol/L (,72 mg/dL) because of non- lesterol, hypertension, alcohol intake, glucose level as a function of time since linearity (Supplementary Fig. 2). In the physical activity, education, and meno- the last meal is shown in Supplementary observational analyses, follow-up began pausal status (for women)dwere asso- Fig. 1 and Supplementary Table 4. Ge- at the first inclusion in a study and ended ciated with glucose levels and with the notype distributions did not deviate from with censoring at the date of death, of an weighted allele score. Hardy-Weinberg expectations (all P . 0.05). event, or of emigration (n 5 610), or on Third, instrumental variable analysis by 10 April 2018 (corresponding to the end two-stage least-squares regression (through Observationally High Glucose Levels of follow-up for the least updated regis- the Stata ivreg2 and ivpois commands) and Risk of Disease try)dwhichever came first. We did not was used in order to estimate the poten- Figure 1 shows the prospective risk of lose track of any individual, as all indi- tial causal association between 1 mmol/L vascular end points by categories of glu- viduals living in Denmark have a Danish higher glucose levels and risk of disease cose level in individuals without diag- Civil Registration System number, which (19,20). The strength of the genetic in- nosed diabetes at baseline. Baseline glucose provides information, updated daily, on strument (i.e., the association between levels were 80% higher (mean 6 SEM emigration and death. The models were genotypes and glucose levels) was con- 8.560.02mmol/L)inthehighestglucose adjusted for sex, birth year, current smok- firmed with the F statistic for a weighted category below the diabetes cutoff and ing, pack-years smoked, BMI, hyperten- allele score of 98, explaining 0.3% of the 186% higher in the highest category above sion, LDL cholesterol, time since last variation in glucose levels (21). the cutoff(13.4 6 0.01mmol/L); the level meal, and menopausal status (for women). Fourth, to validate and test the gen- in the reference category was 4.8 6 0.01 Missing values for covariates (0–2.4% miss- eralizability of the resultsdthat is, how mmol/L) (Fig. 1). Higher glucose levels ing) were imputed from age, sex, and well they represent glucose levels per se were associated with stepwise higher cohort (i.e., whether the individual par- and are able to be reproduced in another risks of all end points, and higher risks ticipated in CCHS or CGPS) by using mul- general populationdwe conducted two- also were observed for glucose levels tivariate imputation. Cohort was adjusted sample MR analyses with a broader ge- within the normal range. in order to accommodate for the time netic instrument based on summary-level Multifactorial adjusted hazard ratios difference between the initiation of the data of 26 genetic variants associated with were 13 (95% CI 8.0–20) for retinopathy, two cohorts (1976 for the CCHS and 2003 glucose levels in the MAGIC participants 2.5 (1.9–3.2) for peripheral neuropathy, for the CGPS), and for potential associ- (n 5 133,010) (13). (For further descrip- 12 (7.6–18) for diabetic nephropathy, and ated changes in patterns of risk behavior, tions of methods of the two-sample MR, 1.3 (1.1–1.6) for PAD in individuals with diagnoses, and treatment options. We see Supplementary Appendix 3.) Genetic nonfasting glucose levels between 8.0 also examined the observational associ- variants with P values ,1 3 1028 were and 11.0 mmol/L. The risks of all end ations between glucose level (on a contin- selected; variants in linkage disequilib- points were even higher in individuals uous scale) and disease byusing restricted rium, defined as an R2 value .0.6, were with a nonfasting glucose level above the cubic splines with seven knots incorpo- excluded. End point data for the variants diabetes cutoff ($11.1 mmol/L) (P for rated into Cox proportional hazards mod- were extracted from the UK Biobank (n 5 trend ,0.001 for all end points) (Fig. 1). elsincluding all individuals (those with and 452,264) (12) and the CKDGen Consor- The known stepwise association between those without diabetes); the population- tium(n5133,413)(14).(Formoredetails higher glucose levels and MI is shown as a wide median glucose level was the ref- about the genetic variants used, see Sup- positive control of the study’s power (P , erence value. plementary Table 2.) 0.001).Restricted cubicspline analyses of Second, to test whether genotypes and Finally, summary-level estimates of the associations between glucose levels the weighted allele score were associated causal effects for the data from the CGPS, (on a continuous scale) and risk of ret- with increased risk of disease, we used CCHS, and consortia populations were inopathy, peripheral neuropathy, diabetic unadjusted Cox proportional hazards re- estimated for each end point through nephropathy, and PAD for all individuals gression (as genotypes generally have a meta-analyses by using the metan com- (i.e., those with and those without di- constant effect throughout life and are mand in Stata. Between-study heteroge- abetes) showed an increasing risk of dis- largely unaffected by confounding fac- neity was assessed by using the I2 statistic. ease with glucose levels increasing from tors), with age as the time scale. Be- If heterogeneity was ,50%, the analyses 5.2 mmol/L (Supplementary Fig. 2). cause an individual’s genotype is constant were performed with a fixed effects model. throughout life, follow-up time began If heterogeneity was .50%, a random Genetically High Glucose Levels and when the national Danish Patient Reg- effects model was chosen (22). Risk of Disease istry was established (1 January 1977) or The selected genetic variants of GCP62/ on the individual’s 20th birthday, which- RESULTS ABCB11, GCK, DGKB, ADCY5, CDKN2A/B, ever came last; follow-up ended as de- A total of 117,193 individuals were in- and TCF7L2 were separately associated scribed for the observational analyses. A cluded from the CCHS and the CGPS. Of with stepwise higher glucose levels, as critical assumption of the MR design is that these, 4,738 had a diagnosis of diabetes was the weighted allele score obtained 4 Impact of Glucose Level on Vascular Disease Diabetes Care

Table 1—Baseline characteristics for all individuals and by random plasma glucose in categories for individuals without diabetes Random plasma glucose in individuals without diabetes, mmol/L (mg/dL) 4.0–4.5 5.5–6.4 6.5–7.9 8.0–11.0 $11.1 P for All individuals (72–98) (99–116) (117–143) (144–199) ($200) trend Individuals 117,193 (100) 75,600 (66) 26,258 (23) 8,162 (7) 2,055 (2) 382 (0.3) Glucose mmol/L 5.1 (4.7–5.7) 4.9 (4.6–5.1) 5.8 (5.5–6.0) 6.9 (6.6–7.3) 8.6 (8.2–9.3) 13.7 (12.0–16.3) ,0.001 mg/dL 92 (85–103) 88 (83–92) 104 (101–108) 124 (119–131) 155 (148–167 246 (216–293) ,0.001 Age, years 58 (48–68) 56 (47–66) 61 (51–70) 63 (53–71) 64 (56–72) 63 (56–71) ,0.001 Female sex 64,580 (55) 43,085 (57) 13,759 (52) 4,167 (51) 1,012 (49) 145 (38) ,0.001 BMI, kg/m2 25.5 (23.1–28.4) 25.3 (23.0–28.0) 25.8 (23.4–28.7) 26.1 (23.5–29.1) 26.8 (24.0–30.1) 29.5 (26.3–32.9) ,0.001 LDL cholesterol mmol/L 3.2 (2.6–3.9) 3.2 (2.6–3.9) 3.3 (2.6–3.9) 3.2 (2.6–3.9) 3.2 (2.3–3.9) 3.5 (2.5–4.3) ,0.001 mg/dL 124 (101–151) 124 (101–151) 124 (101–151) 124 (101–151) 124 (97–151) 135 (97–167) ,0.001 Hypertension 62,852 (54) 37,669 (50) 15,438 (59) 5,015 (61) 1,379 (67) 326 (85) ,0.001 Current smoker 23,140 (20) 14,290 (19) 5,346 (20) 1,799 (22) 538 (26) 111 (29) ,0.001 Pack-yearsa 17 (6–31) 15 (6–30) 18 (7–33) 20 (8–36) 22 (10–39) 28 (15–44) ,0.001 Alcohol intake .7/14 units for women/menb 45,635 (39) 29,180 (39) 10,741 (41) 3,209 (40) 790 (39) 136 (36) ,0.001 Physical activity Low 10,618 (9) 6,188 (8) 2,635 (11) 853 (11) 246 (13) 79 (27) ,0.001 Intermediate 90,665 (81) 59,649 (81) 6,040 (80) 6,040 (80) 1,465 (80) 193 (66) ,0.001 High 11,198 (10) 7,809 (11) 2,156 (9) 637 (8) 124 (7) 19 (7) ,0.001 Education .13 years 21,436 (18) 15,235 (20) 3,998 (15) 1,127 (13) 235 (12) 44 (12) ,0.001 Data are n (%) for categorical variables or median (interquartile range) for continuous variables. Biochemical analyses were performed on samples obtained from participants in a nonfasting state. aPack-years are for only current and former smokers. bOne unit of alcohol corresponds to 12 g.

when those variants were combined: the level and the combined genetic variants. measured reduced kidney function, we mean glucose level was 5% higher in in- Age, sex, menopausal status (for women), performed the analyses for eGFR ,60 dividuals with an allele score above the BMI, hypertension, current smoking sta- mL/min/1.73 m2 and found an observa- 95th percentile (mean 6 SEM 5.56 6 tus, alcohol intake, physical activity, and tional RR of 1.04 (1.02–1.06) and a causal 0.02 mmol/L) than in individuals with an education level were associated with glu- RR of 0.97 (0.84–1.12) per 1 mmol/L higher allele score below the 26th percentile cose level but not with the combined glucose level. Sensitivity analyses of the (mean 6SEM 5.28 6 0.01mmol/L) (Fig.2 genetic variants, indicating that pleiotro- instrumental variables gave similar re- and Supplementary Fig. 3). The effect of pic effects are unlikely through any of the sults when using the seven genetic variants allele frequency and the direction and aforementioned factors (Supplementary in an allele score weighted for their effects magnitude of the effect of each genetic Figs. 4 and 5). on fasting glucose level in participants in variant on random plasma glucose in the the MAGIC instead of their effect on individuals from CCHS and from CGPS Causal Association Between High random plasma glucose in participants werelargelycomparabletothefrequency, Glucose Level and Disease in the CCHS or the CGPS (Supplementary direction, and magnitude of their effects Multifactorial adjusted observational anal- Fig. 6). on fasting plasma glucose in the MAGIC yses and causal analyses of instrumental On the basis of summary-level data for participants(SupplementaryTable5).Cat- variablesforriskestimatesofdiseasepera the 26 genetic variants associated with egories of genetically higher glucose levels 1 mmol/L higher glucose level showed high fasting glucose levels in individuals wereassociatedwithstepwisehigherrisks observational and causal associations be- without diabetes in the MAGIC, combined of retinopathy (P for trend 0.002), periph- tween high glucose levels and risk of ret- with end point data from the UK Biobank eral neuropathy (P 5 0.001), and diabetic inopathy (observational risk ratio [RR] participants into a causal estimate by nephropathy (P 5 0.007) (Fig. 2). The risk 1.32 [95% CI 1.29–1.35]; causal RR 2.01 two-sample MR regression, the inverse- of PAD was higher for individuals in the [95% CI 1.18–3.41]), peripheral neurop- variance weighted (IVW) estimates were highest allele score category than for athy (observational RR 1.16 [1.14–1.19]; similar to the results from the CCHS and those in the lowest (HR 1.19 [95% CI causal RR 2.15 [1.38–3.35]), and diabetic the CGPS combined, that is, an RR of 4.55 1.05–1.35]). Estimates for the known as- nephropathy (observational RR 1.29 [1.25– (95% CI 2.26–9.15) for retinopathy, 1.48 sociation between genetically higher glu- 1.32]; causal RR 1.58 [1.04–2.40]) (Fig. 3). (0.83–2.66) for peripheral neuropathy, cose levels and MI is shown as a positive For PAD, a 1 mmol/L higher glucose level and 1.23 (0.57–2.67) for PAD (Fig. 3). We control of the genetic instrument. had an observational RR of 1.06 (1.04– did not have access to diagnosis codes for 1.09) and a causal RR of 1.19 (0.90–1.58). diabetic nephropathy or eGFR measure- Confounding Factors To investigate whether the causal asso- ments from the UK Biobank. Using summary- We tested whether potentially confound- ciation between high glucose level and level data for the same 26 genetic variants ing factors were associated with glucose diabetic nephropathy also was valid for from the MAGIC combined with end point care.diabetesjournals.org Emanuelsson and Associates 5

Figure 1—Observational prospective risk of retinopathy, peripheral neuropathy, diabetic nephropathy, PAD, and MI as a function of increasing random plasmaglucose(p-glucose)inindividualswithoutdiabetesatbaseline.Individualswithaneventbeforebaselineorwithaglucoselevelbelow4.0mmol/L (n 5 2,103) were excluded from the analyses. N, number HR, 5 hazard ratio,; Δ%, percentage difference in mean glucose level. data for reduced kidney function from causal association between high glucose participants in the MAGIC (13) and the CKDGen Consortium, we found re- level and MI is shown for the CCHS and end point data from the UK Biobank sults similar to those in the CCHS and CGPS combined, the UK Biobank cohort, (12) and the CKDGen Consortium (14). CGPS combined, that is, no evidence of and the CARDIoGRAMplusC4D Consor- Our findings indicate that a high glucose causal association between high glucose tium as a positive control of study power level within the normal range is an level and eGFR ,60 mL/min/1.73 m2 (RR and of the genetic instruments (Fig. 3 and important causal risk factor for micro- 0.98 [95% CI 0.95–1.01]). Combining the Supplementary Fig. 7). vascular disease. Furthermore, our study causal estimates from the CCHS and replicates previous findings showing CGPS with the IWV estimates from CONCLUSIONS that a high glucose level within the the UK Biobank and CKDGen Consortium In this cohort from the Danish general normal range is a causal risk factor for byusingmeta-analyses,a1mmol/Lhigher population, we found observational and MI (1), suggesting that in a general glucose level was associated with a causal causal associations between elevated population, elevated glucose levels over RR of 2.93 (95% CI 1.32–6.50) for reti- glucose level and high risk of retinopa- time may be important in the develop- nopathy, 1.87 (1.32–2.67) for peripheral thy, peripheral neuropathy, and diabetic ment of both micro- and macrovascular neuropathy, 0.98 (0.95–1.01) for eGFR nephropathy. Observationally, we found disease. ,60 mL/min/1.73 m2, and 1.19 (0.90– positive associations between high ran- Previous observational studies have 1.56) for PAD. The corresponding two- dom plasma glucose level and risk of PAD shownthatretinopathy, neuropathy, and sample MR estimates using MR Egger and CKD (defined as eGFR ,60 mL/min/ signs of microvascular dysfunction are regression and weighted median regres- 1.73 m2); however, a causal association prevalent in individuals with sion showed similar results for all end could not be confirmed for PAD and (23–25). Observational findings may be points, and we found no indication of seemed to be refuted for an eGFR ,60 confounded by concomitant cardiome- pleiotropy (all P values $0.10 for the MR mL/min/1.73 m2.Thesefindings were tabolic risk factors such as obesity, hyper- Egger intercept) or bias due to invalid validated by similar results from a two- lipidemia, and hypertension. Our finding instruments (Supplementary Fig. 7). Es- sample MR analysis using summary-level of a stepwise increase in the risk of vas- timates for the known observational and data for fasting glucose levels from cular disease with increasing glucose 6 Impact of Glucose Level on Vascular Disease Diabetes Care

Figure 2—Mean random plasma glucose (p-glucose) level as a function of weighted allele score categories, and the prospective risk of retinopathy, peripheral neuropathy, diabetic nephropathy, PAD, and MI as a function of allele score (by categories of increasing score). Risk of MI is included as a positive control of the genetic instrument. N, number; HR, 5 hazard ratio; Δ%, percentage difference in mean glucose level.

levels within the normoglycemic range or are not likely to be due to glucose low- association between glucose level and higher support the idea that an elevated ering alone but to several beneficial met- CKD (defined as an eGFR ,60 mL/min/ glucose level per se has a causal role in abolic effects. Our findings highlight the 1.73 m2). These findings may be ex- the pathogenesis of microvascular dis- importance of early detection of glycemia plained by the fact that these two end ease, as do levels below the diabetes and screening for prediabetes in asymp- points represent different etiologies of cutoff. This is in line with the general tomatic individuals through the use of kidney disease: Diabetic nephropathy understanding of the natural history of risk assessment toolsdsuch as the one is a microvascular disease with charac- as a continuous process currently provided by the American Di- teristic histopathological changesinduced of declining b-cell function and increas- abetes Association (www.diabetes.org/ by hyperglycemia and is initially charac- ing relative insulindeficiency, leading to a are-you-at-risk/diabetes-risk-test/) (15)d terized by albuminuria, glomerular hy- continuous increase in glucose that is ini- in order to prevent glycemic deteriora- perfiltration, anddthough only in late tiated years before the diabetes threshold tion as early as possible. In addition to stagesdreductions in eGFR (17,31). In is reached (26). Randomized controlled being a risk factor for diabetes and car- contrast, CKD covers a variety of kidney trials have shown that lifestyle changes diovascular disease, our data suggest abnormalities of different etiologies, and and treatment with glucose-lowering that an elevated glucose level below the eGFR is the gold standard but an un- drugs can reduce the progression from diabetes cutoff is an important risk factor specific measure of a decline in kidney prediabetes to diabetes (27–29). Recent for microvascular disease. Screening for function. In many individuals with pre- 30-year follow-up data from a study of retinopathy, neuropathy, diabetic nephrop- diabetes or type 2 diabetes, CKD is 577 Chinese individuals showed that life- athy, and additional risk factors such as probably attributable to a mix of several style interventions in individuals with pre- obesity, hyperlipidemia, and hypertension pathological changes leading to reduced diabetes reduce long-term risks of diabetes, might be indicated in individuals with kidney function, including diabetic ne- a composite of microvascular complica- prediabetes. phropathy, glomerulopathies, previous tions, cardiovascular disease, cardiovas- We found observational and causal episodes of acute kidney disease, and cular mortality, and all-cause mortality associations between glucose level and aging-related nephropathy with micro- (30). The effects of lifestyle intervention diabetic nephropathy, but no causal and macrovascular components (31); care.diabetesjournals.org Emanuelsson and Associates 7

Figure 3—Risk of retinopathy, peripheral neuropathy, diabetic nephropathy, CKD (defined as eGFR ,60 mL/min/1.73 m2), PAD, and MI per 1 mmol/L (18 mg/dL) higher observational and causal glucose levels. Hazard ratios for a 1 mmol/L higher observational glucose level were calculated by using Cox regression in individuals without diabetes at baseline, and RRs for genetically higher glucose levels were derived from instrumental variable analyses (IVAs) of all individualswith available genotypes. Risk was also estimated by using summary-level data from the UK Biobank and the CKDGen Consortium on genetic variants tested in the MAGIC by using MR IVW estimates. The causal estimates from the CCHS and CGPS cohorts and the validation cohorts were combined through meta-analysis. The I2 values denote the percentages of between-study heterogeneity in the meta-analyses. N, number; p-glucose, plasma glucose. hyperglycemia may not be the predom- and thrombosis (33,34). These mechanisms individuals from the general population, inant causal factor. In line with this, the are likely to be similar in coronary ar- the confirmation cohorts used in the renoprotective effects of sodium–glucose teries and large peripheral arteries, and two-sample MR consisted of selected cotransporter 2 inhibitors seen in clinical even though in this study we could not samples that may not be representative intervention trials do not seem to be confirmacausalassociationbetweenglu- of the general population and may thus mediated by lowering glucose, but by cose level and PAD (defined on the basis not be completely comparable (participa- other mechanisms that target common of ICD codes), a previous study of indi- tion rates: CCHS, 61%; CGPS, 43%; UK pathways of CKD progression, regardless viduals without diabetes reported a causal Biobank, 5%) (36). of etiology (31,32). association between fasting glucose level A limitation to our study is that in CCHS Themechanismsbywhichglucosecon- and carotid intima-media thickness (35). and CGPS, glucose was measured from tributes to the pathogenesis of macro- A potential limitation of MR is that the samples obtained from participants in a vascular disease such as MI and PAD are selected genetic variants have pleiotro- nonfasted state, and oral glucose toler- not completely known. Mechanistic stud- pic effects on other risk factors of the ance tests were not performed. Conse- ies suggest that metabolic changes induced diseases studied (21). However, we did quently we could not classify individuals by hyperglycemia, together with increased not find any associations between the according to impaired fasting glucose or and free fatty acids, one-sample MR genetic instrument and impaired glucose tolerance status. In the accelerate the atherosclerotic process age, sex, BMI, smoking, hypertension, LDL observational analyses, estimates were through increased oxidative stress in ar- cholesterol, alcohol intake, physical ac- adjusted for time since the last meal, and terial endothelial cells, the formation of tivity, years of education, or, for women, in the genetic analyses the results were advanced glycation end products, and menopause, and we found no indica- validated by using summary data for nonenzymatic glycation of LDL, apolipo- tions of directional pleiotropy using MR fasting glucose levels. Another limitation proteins, and clotting factors, collectively Egger regression in the two-sample MR. is that eGFR was calculated from a sin- resulting in vasoconstriction, inflammation, Also, even though our aim was to study gle measurement of plasma creatinine, 8 Impact of Glucose Level on Vascular Disease Diabetes Care

which may lead to some misclassification glucose levels within the normal range 2. Sarwar N, Gao P, Seshasai SR, et al.; Emerging of kidney disease. However, the preva- and higher were causally associated with Risk Factors Collaboration. Diabetes mellitus, fast- lence of eGFR ,60 mL/min/1.73 m2 high risks of retinopathy, neuropathy, ing blood glucose concentration, and risk of vascular disease: a collaborative meta-analysis among the CCHS/CGPS population was diabetic nephropathy, and MI. A causal of 102 prospective studies [published correc- ;10%, which corresponds well with association could not be confirmed for tion appears in Lancet 2010;376:958]. Lancet prevalence estimates from other general PAD and seemed to be refuted for eGFR 2010;375:2215–2222 populations (37). Also, we did not have ,60 mL/min/1.73 m2. The findings were 3. Bonadonna R, Cucinotta D, Fedele D, Riccardi G, Tiengo A; Metascreen Writing Committee. The access to albuminuria measurements, validated with similar results by using is a risk indicator of micro- which could have strengthened the val- summary-level data for fasting glucose vascular and macrovascular complications in di- idity of the nephropathy end point. levels from the MAGIC and end point abetes: results from Metascreen, a multicenter Strengths of the study include the ex- data from theUKBiobankandtheCKDGen diabetesclinic-basedsurvey.DiabetesCare2006; – amination of a large number of individ- Consortium. These findings suggest that 29:2701 2707 4. Zoungas S, Arima H, Gerstein HC, et al.; Col- uals from a genetically homogenous general elevated glucose levels should be identi- laborators on Trials of Lowering Glucose (CON- population, access to highly valid data fied as an important risk factor for micro- TROL) group. Effects of intensive glucose control from individual participants, no losses and macrovascular disease in the general on microvascular outcomes in patients with type to follow-up, and the use of MR, which population and that screening for microvas- 2 diabetes: a meta-analysis of individual partic- allowed us to assess potential causal ef- cular disease may be recommended, along ipant data from randomised controlled trials. Lancet Diabetes Endocrinol 2017;5:431–437 fects of high glucose levels on the risk of with screening for additional cardiovascular 5. Davies MJ, D’Alessio DA, Fradkin J, et al. Man- disease, minimizing residual confound- risk factors, in individuals with prediabetes. agement of hyperglycaemia in type 2 diabetes, ing and reverse causation. In addition, 2018. A consensus report by the American Di- the similar results found by using the abetes Association (ADA) and the European As- two-sample MR approach increase the Acknowledgments. The authors thank the staff sociation for the Study of Diabetes (EASD). Diabetologia 2018;61:2461–2498 generalizabilityandvalidityoftheresults, and participants of the Copenhagen General Pop- ulation Study and the Copenhagen City Heart 6. Stehouwer CDA. Microvascular dysfunction and making them reproducible with a differ- Study, and the participants of the Meta-Analyses hyperglycemia: a vicious cycle with widespread ent genetic instrument and in another of Glucose and Insulin-Related Traits Consortium consequences. Diabetes 2018;67:1729–1741 population. The two-sample approach con- (MAGIC), the CKDGen Consortium, the UK Bio- 7. Fedak KM, Bernal A, Capshaw ZA, Gross S. sisted of three separate analyses: the bank, and the CARDIoGRAMplusC4D Consortium Applying the Bradford Hill criteria in the 21st for their generous participation. The authors also century: how data integration has changed causal conventional IVW analysis unadjusted for thank the consortia for making their data publicly inference in molecular epidemiology. Emerg Themes pleiotropy; an analysis that used MR Egger available. Epidemiol 2015;12:14 regression adjusting for directional pleio- Funding. This study was supported by Det Frie 8. Lawlor DA, Harbord RM, Sterne JAC, Timpson N, tropic effects; and a weighted median Forskningsrad˚ (the Danish Council for Indepen- Davey Smith G. Mendelian randomization: using regression accounting for up to 50% of dent Research) (#4183-00171B to M.B.). genesasinstrumentsformaking causal inferences The funder was not involved in the design, in epidemiology. Stat Med 2008;27:1133–1163 information coming from invalid or weak conduct, analyses of the study, or the interpre- 9. Florez JC, Jablonski KA, Bayley N, et al.; Di- instruments (38,39). Risk estimates from tation of the results. abetes Prevention Program Research Group. these three two-sample MR analyses Duality of Interest. B.G.N. reports acting as a TCF7L2 polymorphisms and progression to di- were largely in the same direction and consultant to AstraZeneca, Sanofi, Regeneron, abetes in the Diabetes Prevention Program. N – showed no indication of pleiotropy, con- Akcea Therapeutics, Amgen, Kowa Company, Engl J Med 2006;355:241 250 fi Ltd., Denka Seiken Co., Ltd., Amarin Corporation, 10. Hribal ML, Presta I, Procopio T, et al.; rming the validity of the instrument Novartis, Novo Nordisk, and Silence Therapeu- EUGENE2 Consortium. Glucose tolerance, insulin (38,39). It is still possible, however, that tics; and has received lecture honoraria from sensitivity and insulin release in European non- the genetic instruments influence the Amgen, Kowa Company, Ltd., and Amarin Cor- diabetic carriers of a polymorphism upstream risk of disease through unknown glucose- poration. No other potential conflicts of interest of CDKN2A and CDKN2B. Diabetologia 2011;54: – independent pathways that we could not relevant to this article were reported. 795 802 Author Contributions. F.E. researched data, 11. UKBiobank.UKBiobank:protocolforalarge- control for. An assumption of two-sample performed the analyses, and wrote the manu- scale prospective epidemiological resource [In- MR analyses is that the two samples rep- script. S.M. reviewed and edited the manuscript. ternet], 21 March 2007. Available from www resent similar but nonoverlapping pop- A.T.-H. and B.G.N. collected data and reviewed .ukbiobank.ac.uk/wp-content/uploads/2011/ ulations. Summary data used for the and edited the manuscript. M.B. performed the 11/UK-Biobank-Protocol.pdf. Accessed 9 Feb- analyses and reviewed and edited the manu- two-sample MR analyses in this study ruary 2020 script.All authors discussed the manuscript. M.B. 12. Canela-Xandri O, Rawlik K, Tenesa A. An atlas were from the MAGIC and the CKDGen is the guarantor of this work and, as such, had full of genetic associations in UK Biobank. Nat Genet Consortium, both general population co- access to all the data in the study and takes 2018;50:1593–1599 horts of European ancestry (13,14), and the responsibility for the integrity of the data and the 13. Scott RA, Lagou V, Welch RP, et al.; DIAbetes overlap between two samples was low accuracy of the data analysis. Genetics Replication and Meta-analysis (DIA- Prior Presentation. This study was presented (,4% of individuals participating in both GRAM)Consortium.Large-scaleassociationanal- at the 79th Scientific Sessions of the American yses identify new loci influencing glycemic traits consortia). The UK Biobank is a general Diabetes Association, San Francisco, CA, 7–11 and provide insight into the underlying biological population cohort comprising participants June 2019. pathways. Nat Genet 2012;44:991–1005 of mainly white and British descent, and 14. Pattaro C, Teumer A, Gorski M, et al.; ICBP is not a part of the MAGIC (12,13). The References Consortium; AGEN Consortium; CARDIOGRAM; CCHS and the CGPS were not part of the 1. Benn M, Tybjaerg-Hansen A, McCarthy MI, CHARGe-Heart Failure Group; ECHOGen Con- Jensen GB, Grande P, Nordestgaard BG. Nonfasting sortium. Genetic associations at 53 loci high- MAGIC or the CKDGen Consortium. glucose, ischemic heart disease, and myocardial in- lightcell typesandbiologicalpathwaysrelevant In conclusion, in this cohort from the farction: a Mendelian randomization study. J Am Coll for kidney function. Nat Commun 2016;7: Danish general population, random plasma Cardiol 2012;59:2356–2365 10023 care.diabetesjournals.org Emanuelsson and Associates 9

15. American Diabetes Association. 2. Classifi- 24. Sorensen¨ BM, Houben AJHM, Berendschot 31. Anders H-J, Huber TB, Isermann B, Schiffer cation and diagnosis of diabetes: standards of TTJM, et al. Prediabetes and type 2 diabetes are M. CKD in diabetes: diabetic kidney disease versus medical care in diabetes-2018. Diabetes Care associated with generalized microvascular dys- nondiabetickidney disease.NatRev Nephrol2018; 2018;41(Suppl. 1):S13–S27 function: the Maastricht Study. Circulation 2016; 14:361–377 16. Burgess S, Thompson SG. Use of allele scores 134:1339–1352 32. Perkovic V, Jardine MJ, Neal B, et al.; as instrumental variables for Mendelian random- 25. Azmi S, Ferdousi M, Petropoulos IN, et al. CREDENCE Trial Investigators. Canagliflozin and ization. Int J Epidemiol 2013;42:1134–1144 Corneal confocal microscopy identifies small-fiber renal outcomes in type 2 diabetes and nephrop- 17. KDIGO CKD Work Group. KDIGO 2012 clinical neuropathy in subjects with impaired glucose athy. N Engl J Med 2019;380:2295–2306 practice guideline for the evaluation and man- tolerance who develop type 2 diabetes. Diabetes 33. Shah MS, Brownlee M. Molecular and cel- agement of chronic kidney disease. Kidney Int Care 2015;38:1502–1508 lular mechanisms of cardiovascular disorders in Suppl 2013;3:1–150 26. Phillips LS, Ratner RE, Buse JB, Kahn SE. We diabetes. Circ Res 2016;118:1808–1829 18. Levey AS, Stevens LA, Schmid CH, et al.; CKD- can change the natural history of type 2 diabetes. 34. Soran H, Durrington PN. Susceptibility of LDL EPI (Chronic Kidney Disease Epidemiology Collab- Diabetes Care 2014;37:2668–2676 and its subfractions to glycation. Curr Opin Lipidol oration). A new equation to estimate glomerular 27. Knowler WC, Fowler SE, Hamman RF, et al.; 2011;22:254–261 filtration rate. Ann Intern Med 2009;150:604–612 Diabetes Prevention Program Research Group. 35. Rasmussen-Torvik LJ, Li M, Kao WH, et al. 19. Baum C, Schaffer M, Stillman S. Instrumental 10-year follow-up of diabetes incidence and Association of a fasting glucose genetic risk score variables and GMM: estimation and testing. Stata weight loss in the Diabetes Prevention Program with subclinical atherosclerosis: the Atheroscle- J 2003;3:1–31 Outcomes Study. Lancet 2009;374:1677–1686 rosis Risk in Communities (ARIC) study. Diabetes 20. Nichols A. IVPOIS: stata module to estimate 28. Lindstrom¨ J, Ilanne-Parikka P, Peltonen M, 2011;60:331–335 an instrumental variables Poisson regression via et al.; Finnish Diabetes Prevention Study Group. 36. Munafo` MR, Tilling K, Taylor AE, Evans DM, GMM [article online], 2008. Available from https:// Sustained reduction in the incidence of type 2 Davey Smith G. Collider scope: when selection EconPapers.repec.org/RePEc:boc:bocode:s456890. diabetes by lifestyle intervention: follow-up of bias can substantially influence observed asso- Accessed 9 February 2020 the Finnish Diabetes Prevention Study. Lancet ciations. Int J Epidemiol 2018;47:226–235 21. Benn M, Nordestgaard BG. From genome-wide 2006;368:1673–1679 37. Hill NR, Fatoba ST, Oke JL, et al. Global association studies to Mendelian randomization: 29. Li G, Zhang P, Wang J, et al. The long-term prevalence of chronic kidney disease - a system- novel opportunities for understanding cardiovascu- effect of lifestyle interventions to prevent di- aticreviewandmeta-analysis.PLoS One2016;11: lar disease causality, pathogenesis, prevention, and abetes in the China Da Qing Diabetes Prevention e0158765 treatment. Cardiovasc Res 2018;114:1192–1208 Study: a 20-year follow-up study. Lancet 2008; 38. Bowden J, Davey Smith G, Burgess S. Men- 22. Higgins JPT, Thompson SG, Deeks JJ, Altman 371:1783–1789 delian randomization with invalid instruments: DG. Measuring inconsistency in meta-analyses. 30. Gong Q, Zhang P, Wang J, et al.; Da Qing effect estimation and bias detection through BMJ 2003;327:557–560 Diabetes Prevention Study Group. Morbidity and Egger regression. Int J Epidemiol 2015;44:512– 23. Abdul-Ghani M, DeFronzo RA, Jayyousi A. mortality after lifestyle intervention for people 525 Prediabetes and risk of diabetes and associated with impaired glucose tolerance: 30-year results 39. Burgess S, Butterworth A, Thompson SG. complications: impaired fasting glucose versus of the Da Qing Diabetes Prevention Outcome Mendelian randomization analysis with multiple impaired glucose tolerance: does it matter? Curr Study. Lancet Diabetes Endocrinol 2019;7:452– genetic variants using summarized data. Genet Opin Clin Nutr Metab Care 2016;19:394–399 461 Epidemiol 2013;37:658–665