Profiles of Glucose Metabolism in Different Prediabetes Phenotypes, Classified by Fasting Glycemia, 2-Hour OGTT, Glycated Hemoglobin, and 1-Hour OGTT: An IMI DIRECT Study

The current prediabetes conditions include impaired fasting glucose (IFG), impaired glucose tolerance (IGT), and HbA_1c indicative of prediabetes (IA1c) (1). In some studies investigators found that IFG and IGT differ in the mechanisms involved in glucose homeostasis (2); they summarized that, though both individuals with IFG and IGT show reduction in early-phase insulin secretion, those with IGT also have impaired late-phase insulin secretion. Furthermore, individuals with IGT have marked peripheral insulin resistance with only mild hepatic insulin resistance, whereas those with IFG show the opposite condition. However, few studies considered all the prediabetes groups and their combinations, and they suffered from several limitations, especially lack of longitudinal data (3–6).

In this study, we investigated the differences in the main parameters of glucose metabolism in all categories of prediabetes, i.e., IFG, IGT, IA1c, and their combinations. Furthermore, we considered 1-h oral glucose tolerance test (OGTT) glycemia, since increasing consensus is emerging for this criterion to characterize glucose tolerance (7). We also investigated differences between groups in the incidence of type 2 diabetes onset after 48 months. Finally, we briefly investigated reversal to normal glucose tolerance (NGT).

We used data from the Innovative Medicines Initiative (IMI) Diabetes Research on Patient Stratification (DIRECT) European multicenter project, aimed to validate biomarkers of glycemic deterioration before and after type 2 diabetes onset (ClinicalTrials.gov identifier NCT03814915) (8) The present analysis considers a cohort of European adults without diabetes, with focus on data collected at baseline (month 0) and month 48. A screening tool was used to identify, from previous cohort studies, at-risk participants to be recruited into the new study (8). Inclusion criteria were as follows: 1) no treatment with insulin-sensitizing, glucose-lowering, or other antidiabetes drugs; 2) fasting capillary blood glucose <10 mmol/L at baseline; 3) White European ethnicity; and 4) age ≥35 and <75 years. Exclusion criteria were 1) diagnosed diabetes of any type; 2) for women, pregnancy, lactation, or plans to conceive within the study period; and 3) any significant medical reason for exclusion as determined by the investigators. All participants provided written informed consent, and the study protocol was approved by the regional research ethics review boards. The research conformed to the ethics principles for medical research involving human participants outlined in the Declaration of Helsinki.

Participants underwent a 75-g OGTT at baseline and at 48 months, with measurement of glucose, insulin, and C-peptide at 0, 15, 30, 45, 60, 90, and 120 min. Plasma glucose was analyzed with an enzymatic method and photometric measurement. Plasma insulin and C-peptide were analyzed with chemiluminometric immunoassay. HbA_1c was measured by liquid chromatography. Assays were carried out centrally at the University of Eastern Finland for glucose, insulin, and C-peptide (within- and between-run coefficients of variation ≤6.6%) and the University of Exeter (Exeter, U.K.) for HbA_1c (coefficients of variation ≤3%). Additional details have previously been reported (8).

Based on the definition of prediabetes of the American Diabetes Association (ADA) (1) we classified the participants as having IFG (fasting glucose ≥5.6 and ≤6.9 mmol/L), IGT (2-h glucose ≥7.8 and ≤11.0 mmol/L), or IA1c (HbA_1c ≥39 and ≤47 mmol/mol). We stratified the participants in groups having a single defect (isolated IFG, IGT, or IA1c), two defects (IFG+IGT, IFG+IA1c, or IGT+IA1c), or all three defects (IFG+IGT+IA1c). We also considered participants with NGT (fasting glucose <5.6 and 2-h glucose <7.8 mmol/L and HbA_1c <39 mmol/mol). In a separate analysis, we also considered participants having the single defect of 1-h hyperglycemia (7) (I1hG) (1-h glycemia ≥8.6 mmol/L), thus without the defects of fasting glucose, 2-h glucose, and HbA_1c (i.e., <5.6 mmol/L, <7.8 mmol/L, and <39 mmol/mol, respectively). In this analysis, the IFG, IGT, IA1c, and NGT groups were redefined by exclusion from each group of the participants with 1-h hyperglycemia.

From the 2,127 participants initially included, 16 were excluded due to lack of data relevant for the analyses; thus, 2,111 participants were studied. The 1,691 participants who completed the final examination at 48 months were analyzed for determination of the incidence of type 2 diabetes, diagnosed according to at least one of the American Diabetes Association criteria (1) or based on records of clinical diagnosis or use of antidiabetes medications.

β-Cell function (BCF) was assessed by mathematical modeling and quantified by glucose sensitivity (G_SENS) (slope of relationship between insulin secretion and glucose concentration), rate sensitivity (R_SENS) (index of early secretion), insulin secretion at reference glucose of 6 mmol/L (rounded mean basal glucose in all participants) (ISR_r), potentiation factor ratio (PFR) (index of OGTT insulin secretion potentiation), and basal (ISR_b) and total (ISR_t) insulin secretion (9). Insulin resistance was estimated at fasting by the homeostasis model assessment index of insulin resistance (HOMA-IR) (10) and insulin sensitivity from the OGTT by predicted M (PREDIM) (11), a surrogate of the clamp M value. Insulin clearance (CL_ins) was obtained as the ratio between the area under the curve of total insulin secretion and that of plasma insulin (12).

Normality of parameter distributions was tested with the Shapiro-Wilk test. In case of skewed distributions, values were logarithmically transformed. Differences among groups in parameter means were assessed with one-way ANCOVA, with adjustment for sex, age, and BMI. Pairwise comparisons were also performed. In case of inhomogeneity of parameter variances, assessed with the Levene test, heteroscedasticity was addressed by generalized least squares allowing separate variances per group. Moreover, we used Tukey honestly significant difference to adjust for multiple statistical testing. Logistic regression analysis was used to assess differences and odds ratios (ORs) for type 2 diabetes incidence in the studied groups (or similar analyses), with adjustment for sex, age, and BMI. Parameter changes between baseline and follow-up were assessed by paired t test, following logarithmic transformation for skewed distributions and assuming inequality of variances if appropriate. Difference in sex distribution among groups was assessed by χ² test. Values are presented as mean ± SD unless otherwise specified. Two-sided P < 0.05 was considered statistically significant.

Due to the type of consent provided by study participants and the ethics approvals for this study, individual-level data from IMI DIRECT cohorts cannot be transferred from the centralized IMI DIRECT repository. Requests for access to IMI DIRECT data, including those presented here, can be made to [email protected]. Requestors will be provided with information and assistance on how data can be accessed via the DIRECT Computerome secure analysis platform following submission of appropriate documentation.

We first compared participants with single defect pooled (1DEF) with the NGT group (Table 1). 1DEF were slightly younger (P < 0.04) but had higher BMI (P < 0.0001). As expected, glycemia and HbA_1c were higher (P < 0.0001).

1DEF had worse insulin sensitivity both at fasting and during the OGTT, as assessed with HOMA-IR and PREDIM (P < 0.0001). 1DEF also showed higher insulin secretion, both at fasting, ISR_b, and total during the OGTT, ISR_t (P < 0.0001). In contrast, ISR_r was lower (P < 0.0001). β-Cell G_SENS was impaired (P < 0.002). CL_ins was lower (P < 0.0001). Supplementary Fig. 1A shows the model-determined relationship between insulin secretion and glucose concentration (i.e., the dose response, whose average slope is G_SENS) for NGT and 1DEF, as well as for other groups, as outlined below.

In comparisons of double defect participants pooled (2DEF) with 1DEF (Table 1), 2DEF were slightly older (P < 0.05). Both glycemia and HbA_1c were higher (P < 0.0001).

2DEF had worse insulin sensitivity, both HOMA-IR and PREDIM (P < 0.003). 2DEF also showed higher ISR_b and ISR_t but lower ISR_r (P < 0.003). G_SENS was impaired (P < 0.0001). Slight impairment was found for the PFR (P < 0.02). β-Cell dose response for 2DEF is reported in Supplementary Fig. 1A.

In comparisons of the group with all three defects (3DEF) (i.e., IFG+IGT+IA1c) with 2DEF (Table 1), 3DEF were slightly older (P < 0.05) and had higher BMI (P < 0.0006). Glycemia and HbA_1c were higher (P < 0.0002).

3DEF had both worse HOMA-IR and PREDIM (P < 0.002). 3DEF also showed higher ISR_b and ISR_t and lower ISR_r (P < 0.02). G_SENS and PFR were lower (P < 0.03). β-Cell dose response for 3DEF is reported in Supplementary Fig. 1A.

The percentage of participants with single defect, compared with that of participants with double or triple defect, is displayed in Supplementary Fig. 2A. In comparison of the single defect groups (Table 2), those with IFG were slightly younger (P < 0.02) than those with IGT and those with IA1c. Glycemia was typically lower for the IA1c group, intermediate for IFG, and higher for IGT (P < 0.0001).

The IFG group showed worse HOMA-IR (P < 0.0001) compared with IGT and IA1c, whereas PREDIM was lower in IGT, intermediate in IFG, and higher in IA1c (P < 0.0001). Those with IFG had higher ISR_b compared with IA1c (P < 0.002), whereas ISR_t was higher in IGT (P < 0.001). The IGT group also had lower ISR_r compared with IFG and IA1c (P < 0.002). In contrast, in the IA1c group, ISR_r was higher than in both IGT and IFG (P < 0.0001). PFR was different in the three groups, being lower in IGT, intermediate in IA1c, and higher in IFG (P < 0.0003). A summary of the differences among groups for the main parameters of insulin sensitivity/resistance and insulin secretion/BCF is reported in Fig. 1A.

The percentage of participants with double defect is displayed in Supplementary Fig. 2A. In comparisons of the double defect groups (Table 2), IGT+IA1c were somehow older (P < 0.008) than the other two groups. Expectedly, the two groups including IFG had higher fasting glycemia than IGT+IA1c (P < 0.0001), whereas the two groups including IGT had higher 2-h glycemia than IFG+IA1c (P < 0.0001); 1-h glycemia was lower in IFG+IA1c than in the other two groups (P < 0.003). Mean glycemia was lower in IFG+IA1c, intermediate in IGT+IA1c, and higher in IFG+IGT (P < 0.004), whereas HbA_1c was higher in the two groups including IA1c (P < 0.0001).

Insulin sensitivity during the OGTT was lower in IFG+IGT, intermediate in IGT+IA1c, and higher in IFG+IA1c (PREDIM, P < 0.03), whereas HOMA-IR was only slightly higher in IFG+IGT than in IGT+IA1c (P < 0.05). ISR_t was higher in IFG+IGT compared with IFG+IA1c (P < 0.0001). ISR_r was lower in IFG+IGT than the other two groups (P < 0.005). PFR was different in the three groups, being lower in IGT+IA1c, intermediate in IFG+IGT, and higher in IFG+IA1c (P < 0.0001 for the difference between IFG+IA1c and the other two groups; P < 0.05 for the difference between IFG+IGT and IGT+IA1c). A summary of the differences among groups for the main parameters of insulin sensitivity/resistance and insulin secretion/BCF can be found in Fig. 1B.

The percentage of participants with 1-h hyperglycemia, compared with that of participants with a traditional single defect, is displayed in Supplementary Fig. 2B. Of note, the number of participants in the groups is different from that in the previous analyses, as we have now considered one additional criterion. We first compared I1hG with NGT (Table 3). Expectedly, glycemia was higher in I1hG (P < 0.0001).

I1hG had worse PREDIM (P < 0.0001). I1hG also showed higher ISR_t and lower ISR_r (P < 0.0001). G_SENS was impaired in I1hG (P < 0.0001). CL_ins was lower (P < 0.02).

We then compared I1hG with the other single defect groups (Table 3) excluding IGT, due to low number of participants (N = 5). I1hG were younger than IA1c (P < 0.005). Glycemia was higher in I1hG (P < 0.0002) than in the other two groups, except for fasting glycemia compared with IFG, as expected; HbA_1c was obviously lower than in IA1c (P < 0.0001) but was similar to that in IFG.

I1hG had better HOMA-IR compared with IFG (P < 0.0001) but lower PREDIM compared with IA1c (P < 0.0009). I1hG also showed higher ISR_t (P < 0.0001) but lower ISR_b compared with IFG (P < 0.0001). Also, in I1hG, ISR_r was lower (P < 0.0001). G_SENS was impaired (P < 0.0008) in I1hG, who also showed lower PFR compared with that in the IFG group (P < 0.005). The β-cell dose response for the groups of this analysis is reported in Supplementary Fig. 1B. A summary of the differences among groups for the main parameters of insulin sensitivity/resistance and insulin secretion/BCF is reported in Fig. 2.

We evaluated how many participants developed type 2 diabetes by the 48-month follow-up visit. Among the 1,691 participants studied, the percentage who developed diabetes in each group is shown in Fig. 3. According to logistic regression analysis, 1DEF showed higher percentage of diabetes compared with NGT (P < 0.0002), with OR much higher than 1 (OR 6.23, 95% CI 2.45–15.87). However, among the three groups with single defect there were no differences in the percentage of participants developing diabetes (P ≥ 0.64).

Similarly, in comparisons of 2DEF with 1DEF, 2DEF showed a higher percentage of participants with diabetes (P < 0.0001, OR 3.03, 95% CI 1.99–4.61). However, there were no differences in the percentage of participants with diabetes among the three groups with double defect (P ≥ 0.59). In comparison of 3DEF with 2DEF, the former had a higher percentage of participants developing diabetes (P < 0.008, OR 2.18, 95% CI 1.23–3.88).

Of note, when analyzing progression to the triple prediabetes defect rather than to overt diabetes, we found higher incidence in IGT than in IFG and IA1c (9 of 43 vs. 42 of 538 and 11 of 157 participants; P < 0.007, OR 3.10, 95% CI 1.37–7.04, and P < 0.008, OR 3.84, 95% CI 1.43–10.31, respectively), as well as in IGT+IA1c than in IFG+IA1c (P < 0.02, OR 4.46, 95% CI 1.31–15.12).

In the analysis including I1hG, 2 of 381 participants developed diabetes in NGT, 7 of 231 in IFG, 4 of 86 in IA1c, and 3 of 150 in I1hG. There were no differences in the percentage of participants developing diabetes in I1hG compared with the other two groups with single defect, i.e., IFG and IA1c (P > 0.13). However, single defect groups pooled (i.e., IFG, IA1c, I1hG grouped) showed higher percentage of diabetes onset compared with NGT (P < 0.03, OR 5.71, 95% CI 1.29–25.29), consistent with previous findings.

For participants who completed the follow-up, we also compared the parameters value at baseline and 48 months. All groups, including NGT, showed deterioration of several metabolic parameters (Table 4 and Table 5).

We also investigated the incidence of reversal to NGT at follow-up. In the groups with single defect, the fraction of participants reverting to NGT was 26 of 538 in IFG, 4 of 43 in IGT, and 2 of 157 in IA1c. The percentage was lower in IA1c than in IFG and IGT (P < 0.03, OR 0.18, 95% CI 0.04–0.80 vs. IFG and P < 0.01, OR 0.08, 95% CI 0.01–0.54 vs. IGT), whereas there was no difference between IFG and IGT (P > 0.34). In groups with double or triple defect, reversal was negligible (2 of 421 participants in total). In the analysis with 1-h hyperglycemia, there was no difference between I1hG and the other groups with single defect (P > 0.08).

In this study, we investigated the differences in the main parameters of glucose metabolism in the prediabetes categories according to all established criteria, i.e., defect in fasting glycemia, 2-h OGTT glycemia, and glycated hemoglobin (1). We also investigated the differences among groups in the incidence of type 2 diabetes. To our knowledge, this is the first study presenting the analysis of glucose metabolism profiles in groups identified by all prediabetes criteria, in isolation and combination. Thus, even for the widely studied IFG and IGT populations, previous investigations typically did not analyze the “pure,” single metabolic defects (at basal or at 2 h), since they rarely accounted for the third possible defect, i.e., IA1c. Furthermore, none of the previous studies comparing the different prediabetes defects reported longitudinal data.

Among the groups with isolated prediabetes defect, we found differences in the type or degree of impairment for both insulin sensitivity and BCF. Similar results were found for the groups with double defect. In line with the concepts of precision medicine in diabetes (13), our findings indicate the potential clinical benefit of treating each category of prediabetes with optimized strategies, the success of which in preventing or delaying type 2 diabetes appears enhanced with interventions designed to correct the underlying pathophysiological disturbances (14).

In comparisons of the groups with isolated defect, in IFG and IGT our findings confirmed the known notions on fasting insulin resistance in IFG (partly reflecting hepatic insulin resistance [15]) and OGTT insulin resistance in IGT (2). In IGT we also found lower ISR_r and lower PFR, which is related to the enhancing incretin effect on insulin secretion (16), though this cannot be investigated in detail in this analysis. In fact, incretin effect alterations were observed in IGT (16,17). In summary, IGT appears to be the worst phenotype among the three phenotypes with isolated defect.

In IA1c, we found fasting insulin resistance similar to that of IGT, but lower than that of IFG, whereas OGTT insulin sensitivity was higher than for both IFG and IGT. IA1c also had ISR_r higher than both IFG and IGT. Glucose and rate sensitivities were similar to those of IFG and IGT, whereas PFR was somehow lower than in IFG but higher than in IGT. Thus, among the three groups with isolated defect, IA1c showed overall less severe impairment in glucose metabolism. Notably, to our knowledge only one study reported detailed (i.e., model-derived) information on BCF in isolated HbA_1c-based prediabetes (18), and comparison with other groups with prediabetes was limited.

In comparisons of the groups with two defects, our findings are consistent with those in the groups with isolated defect, where the metabolic impairment appears more severe in IGT and, in contrast, less severe in IA1c. Thus, IFG+IGT was the worst phenotype, with more severe impairment in both fasting and OGTT insulin sensitivity and partly in BCF.

One-hour hyperglycemia has been proposed as a possible additional marker indicating abnormal glucose metabolism (7,19–22). In our analysis, I1hG showed differences in both insulin sensitivity and BCF compared with the other groups with isolated defect, though comparison with IGT was not possible. Interestingly, I1hG showed impairment in both G_SENS (typically the most important BCF parameter) and potentiation factor, as well as in ISR_r. Thus, 1-h hyperglycemia may identify a prediabetes phenotype with specific metabolic defects (especially, possibly, greater β-cell dysfunction). To our knowledge, no previous studies compared isolated 1-h hyperglycemia with isolated IFG and isolated IA1c.

We also studied the incidence of type 2 diabetes within 4 years from baseline. We did not find differences in the diabetes incidence among the groups with isolated defect (despite our findings suggesting more severe metabolic impairment in IGT) or among the double defect groups. These findings, which partly differ from those of some previous studies in prediabetes (23), may be due to the follow-up duration, which could have been insufficient to disclose possible differences in diabetes incidence among groups with equal number of defects. This is in fact suggested by the observation that, when comparing IFG, IGT, and IA1c in terms of progression to the triple prediabetes defect rather than to overt diabetes, we found higher incidence for IGT compared with both IFG and IA1c and for IGT+IA1c compared with IFG+IA1c. Thus, it can be hypothesized that with longer follow-up duration some differences may emerge in diabetes incidence, among both the groups with isolated defect and those with double defect (with IGT phenotype possibly more prone to development of diabetes, both as isolated defect and in combination with a second defect). On the other hand, 4 years of follow-up was sufficient to reveal differences in diabetes incidence for progressively higher number of prediabetes defects, i.e., from NGT to triple defect. In the context of precision diagnostics (13), this suggests that people known to be at risk for type 2 diabetes should ideally be tested to disclose the possible presence of all three prediabetes defects, as this appears relevant for determination of the actual risk of developing diabetes.

Recent studies suggest that CL_ins is an independent process that can adapt to the metabolic demand to maintain glucose homeostasis (24). Of note, one study reported that both increase in insulin secretion and decrease in CL_ins may compensate for insulin resistance, but CL_ins decrease may be the first mechanism providing adaptation to insulin resistance (25). In our study, we found that CL_ins was similar among groups with equal number of defects, but it showed a tendency to progressive decline, ranging from NGT to triple defect. This appears consistent with what was reported in some studies (24,26), though other studies reported different findings (27,28). The RISE Consortium (27) reported no difference in CL_ins between participants with and without diabetes. Ohashi et al. (28) found an increased hepatic component in type 2 diabetes but decreased extrahepatic component. In a previous study, we found increased CL_ins in women with former gestational diabetes mellitus progressing to type 2 diabetes compared with women remaining diabetes free (12), possibly explained by the role of the SLC30A8 gene (29,30).

Few previous studies analyzed the glucose metabolism profile, including both insulin sensitivity and BCF, in several different groups with prediabetes. In one study (3), participants were stratified into groups of isolated IFG, IGT, and IA1c and further into two groups, with HbA_1c combined with IFG and IGT, but IFG+IGT and triple defect were not considered. Nonetheless, similar to our findings, fasting insulin resistance was higher in IFG and in IFG plus IA1c. BCF results of this study were limited by the lack of C-peptide. This may explain the partial disagreement with our findings, as the worse IGT condition, in BCF, was observed in comparisons with IA1c but not IFG. Furthermore, 1-h glycemia defect was not considered and longitudinal data were not provided. Similar limitations hold for other studies (4,5). In our previous study (6), due to limitations in the data set, we analyzed only IFG and IGT combined, plus one group with added 1-h hyperglycemia and another group with additional HbA_1c defect. In agreement with the present findings, we found progressive deterioration for both insulin sensitivity and BCF for increasing number of defects. However, similarly to the other previously published studies (3–5), C-peptide was not measured and longitudinal data were not available. Other studies reported separate analyses for some prediabetes categories (31–36). However, the focus was typically not on the assessment of glucose metabolism; thus, the analysis of insulin sensitivity and BCF was limited or absent. Another study focused on the definition of prediabetes phenotypes with different metabolic abnormalities and risk for type 2 diabetes (37). Specifically, several variables were considered (glycemia, insulin sensitivity and secretion, lipids, tissue fat content, anthropometry, polygenic diabetes risk score), yielding the definition of six prediabetes phenotypes at different risk for type 2 diabetes. However, direct comparison with our findings is difficult, due to the peculiar definition of the different prediabetes phenotypes (37).

Some studies, e.g., 38–43, focused on reversal from prediabetes to NGT. Interestingly, in our study we found somehow lower reversal in IA1c compared with IFG and IGT, which may be partly due to greater stability of HbA_1c compared with glycemia. This, however, requires further investigations.

Our study has some limitations. The prevalence of the different prediabetes conditions is likely not representative of that among the general population, as it appears from the disparity in IFG and IGT proportions. This is due to the recruitment process (see the study inclusion criteria). However, the size of the study allowed recruitment of a sufficient number of participants even in the less represented prediabetes categories, thus allowing appropriate analyses. Another limitation concerning diabetes incidence may be the duration of the follow-up.

In conclusion, we investigated the differences in several parameters of glucose metabolism in all groups with prediabetes, and we assessed the incidence at 48 months of type 2 diabetes onset in each group. Heterogeneity in the level of impairment of the metabolic parameters suggests that the different prediabetes phenotypes may benefit from specific treatment approaches. Furthermore, our findings indicate that people known to be at risk for diabetes should be tested to disclose the possible presence of all three prediabetes defects, as this appears relevant for determination of the actual risk of developing the disease.

Acknowledgments. The authors thank the participants across all IMI DIRECT study centers for their contributions to the study. The authors also thank the staff involved in the design, implementation, and conduct of the study. A complete list of members of the IMI DIRECT writing group can be found in Supplementary Material.

Funding and Duality of Interest. The work leading to this publication has received support from the Innovative Medicines Initiative Joint Undertaking under grant agreement no. 115317, resources of which are composed of financial contribution from the European Union's Seventh Framework Programme (FP7/2007-2013) and European Federation of Pharmaceutical Industries and Associations (EFPIA) companies’ in kind contribution. R.W.K. received consulting fees from Novo Nordisk and was also funded by a STAR Award Novo Nordisk co-financed PhD fellowship and a Novo Nordisk Foundation postdoctoral fellowship (NNF18OC0031650). P.W.F. has received research funding from Boehringer Ingelheim, Eli Lilly, Janssen, Novo Nordisk A/S, Sanofi, and Servier; received consulting fees from Eli Lilly, Novo Nordisk, and Zoe Global; and has stock options in Zoe Global. H.R. is an employee of Boehringer Ingelheim and a shareholder of Sanofi. M.I.M. was a Wellcome Investigator (grants: 090532, 098381, 106130, 203141, 212259) and a National Institute for Health Research (NIHR) Senior Investigator; has served on advisory panels for Pfizer, Novo Nordisk, and Zoe Global; has received honoraria from Merck, Pfizer, Novo Nordisk, and Eli Lilly and research funding from AbbVie, AstraZeneca, Boehringer Ingelheim, Eli Lilly, Janssen, Merck, Novo Nordisk, Pfizer, Roche, Sanofi, Servier, and Takeda; and, as of June 2019, is an employee of Genentech and a holder of Roche stock. M.R. is an employee of Novo Nordisk. As of January 2020, A.Mah. is an employee of Genentech and a holder of Roche stock. No other potential conflicts of interest relevant to this article reported.

M.I.M. declares that the views expressed in this article are those of the authors and not necessarily those of the National Health Service, the NIHR, or the Department of Health.

Author Contributions. A.T. designed the analysis and analyzed the data. E.G. reviewed the modeling analysis. C.S.G. supervised the statistical analysis. A.Mar. supervised the whole analysis. A.T., E.G., C.S.G., and A.Mar. interpreted the results. A.T. wrote the manuscript. E.G., C.G., and A.Mar. reviewed the manuscript. All authors were involved in the DIRECT study at different levels and contributed to the production, release, and management of the data analyzed here. All authors approved the final version of the manuscript. A.T. and A.Mar. are the guarantors of this work and, as such, had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.