Type 2 Diabetes Subgroups, Risk for Complications, and Differential Effects Due to an Intensive Lifestyle Intervention

In 2013, the primary findings for the Action for Health in Diabetes (Look AHEAD) trial were first reported and demonstrated that a multidomain intensive lifestyle intervention (ILI) did not reduce the rate of cardiovascular disease (CVD) events compared with a diabetes support and education (DSE) control (1). Although the intervention did not demonstrate benefit for CVD events, it did result in reduced weight and improved physical fitness, glucose control, and CVD risk factors among participants (2,3). Weight loss through improvements in diet and physical activity is consistently observed with beneficial effects on various health outcomes and remains recommended for overweight and obese individuals with type 2 diabetes mellitus (T2D) by the American Diabetes Association (4).

In the time since the main results of the Look AHEAD trial were published, our understanding of the heterogeneity of T2D has expanded, which may help give context to the original primary findings of the trial. Multiple studies have investigated whether diabetes, predominantly T2D, can be partitioned into subgroups that differ in underlying genetic risk, clinical profile, and risk for subsequent complications (5–11). This growing body of research suggests that although excess weight and hyperglycemia may be common factors among individuals with T2D, disease etiology and progression may require targeted prevention and treatment strategies.

Given the potential for an individual’s diabetes subgroup to affect their response to pharmacotherapy, we set forth to characterize diabetes subgroups among Look AHEAD participants and evaluate whether the lifestyle intervention was associated with differential rates of CVD events among these diabetes subgroups (11). Our primary objectives were to characterize diabetes subgroups on the basis of the composition of the Look AHEAD sample and determine whether the ILI was associated with differential risk for the primary and secondary cardiovascular outcomes by diabetes subgroup. Previous work in Look AHEAD supported heterogeneity in intervention response based on prior history of CVD and risk factor profiles; thus, we hypothesized that the ILI would be associated with differential risk for all outcomes by diabetes subgroups (1,12–15). With this information, we may be able to develop intervention strategies that are targeted to individuals within a particular diabetes subgroup.

The study methods of the Look AHEAD trial have been described (16). Briefly, Look AHEAD was a randomized controlled trial among individuals with T2D and overweight or obesity. From August 2001 through April 2004, 5,145 participants were recruited at 16 clinical sites across the United States. Eligibility criteria for Look AHEAD included being aged 45–75 years; having T2D verified by the use of glucose-lowering medication, a physician’s report, or glucose levels; and a BMI of ≥25.0 kg/m² (≥27.0 kg/m² if taking insulin). Individuals were excluded from study participation for any of the following: glycated hemoglobin level (HbA_1c) of >11% (>97 mmol/mol); systolic blood pressure (SBP) ≥160 mmHg; diastolic blood pressure ≥100 mmHg; triglyceride level ≥6.77 mmol/L; inability to complete a valid maximal exercise test; and lack of an established relationship with a primary care provider. Participants provided informed consent and local institutional review boards approved the study protocol.

Study participants were randomly assigned, within clinical site and with equal probability, to receive either a multidomain ILI or a control condition of DSE (16–18). The goal of the ILI was for individuals to achieve and maintain weight loss of ≥10% through decreased caloric intake and increased physical activity. Participants who were randomly assigned to DSE were invited to three group educational sessions each year in the first 4 years of the trial, and then to one annual session thereafter. All adjustments to medications were made by the patient’s health care provider, with the exception of temporary changes in glucose-lowering medications made by study staff to reduce the risk of hypoglycemia in the ILI group. Although participants were not blinded to intervention allocation, nonintervention study staff and investigators were unaware of intervention status.

At baseline and annual study clinic visits, staff members certified in assessment procedures measured participants’ height, weight, waist circumference, and blood pressure. With participants in light clothing, height and weight were measured in duplicate using a standard stadiometer and digital scale, respectively. Resting seated blood pressure was measured in duplicate with a Dinamap Monitor Pro100 automated device. Fasting blood was collected and aliquoted, and samples were shipped on dry ice to the Look AHEAD Central Biochemistry Laboratory (Northwest Lipid Research Laboratories, Seattle, WA) where analyses were performed. HbA_1c and HDL cholesterol were measured according to standardized laboratory procedures. LDL cholesterol levels were estimated using the Friedewald equation (19). Interviewer-administered questionnaires were used to collect information on participant medical history, employment, education, family income, prior pregnancies, smoking, prescription medications, alcohol use, and family medical history.

We used the trial’s prespecified primary and secondary end points (1,16). The primary end point was the first occurrence of a composite cardiovascular outcome. The original composite outcome included death from cardiovascular causes, nonfatal myocardial infarction, and nonfatal stroke over an anticipated maximal follow-up period of 11.5 years. When the Data and Safety Monitoring Board assessed the study progress at the 2-year mark, the primary-event rate in the control group was lower than expected (20). At that point, hospitalization for angina was added to the primary outcome and follow-up was extended to a maximum of 13.5 years. Three composite, secondary cardiovascular outcomes were also assessed with the original trial objectives: 1) death from cardiovascular causes, nonfatal myocardial infarction, or nonfatal stroke (original primary outcome); 2) death from any cause, myocardial infarction, stroke, or hospitalization for angina; and 3) death from any cause, myocardial infarction, stroke, hospitalization for angina, coronary-artery bypass grafting (CABG), percutaneous coronary intervention, hospitalization for heart failure, carotid endarterectomy, or peripheral vascular disease. At annual visits and 6-month phone calls, participants were queried by staff masked to intervention assignment about all medical events and hospitalizations. These were augmented with searches of national databases on deaths. Hospital records were obtained for potential CVD events and adjudicated according to standard criteria by reviewers masked to intervention assignment (1).

We applied k-means clustering to characterize diabetes subgroups at the time of randomization, similar to prior studies on diabetes subgroups (6,8,9). Clustering variables included age at diabetes diagnosis, BMI, waist circumference, and HbA_1c. We chose these variables for their importance in predicting T2D development or monitoring diabetes progression, availability for ascertainment in observational and clinical settings, and overlap with prior research on diabetes subgroups (6,8,9). We used two-step, fully conditional specification imputation methods to generate missing data for age at diagnosis (n = 32 observations) and waist circumference (n = 6 observations) (21,22). Before the clustering step, we regressed sex on the clustering variables, separately, and used the residuals as the value for each clustering variable in the k-means models. We calculated Jaccard (range: 0–1; 1 is stable cluster assignment) and Silhouette index (range: −1 to 1; 1 is well matched within cluster) values to assess cluster assignment, stability, and similarity within cluster. The Jaccard and Silhouette metrics supported cluster numbers of three (Jaccard: 0.79; Silhouette: 0.38) or four (Jaccard: 0.77; Silhouette: 0.35) as those that would best fit the data, and we ultimately selected four clusters (subgroups) because this permitted better separation between cluster profiles for each of the clustering characteristics.

After characterizing diabetes subgroups, we assessed the randomization allocation, demographic composition, and clinical characteristics of each subgroup. We assessed whether the ILI was associated with differential risk for the primary and secondary study end points according to diabetes subgroup (four-level categorical variable). To do so, we used Cox proportional hazards regression and included main effects for intervention, study site, history of CVD, and diabetes subgroup and a product term of intervention and diabetes subgroup. A two-sided P value for this product term <0.05 was considered evidence to support an interaction.

We started with assessment of the primary outcome and observed evidence for an interaction between intervention and diabetes subgroup, which supported the use of models stratified by diabetes subgroup. We were able to maintain randomization with these stratified models. We used these stratified models to estimate hazard ratios (HRs) and 95% confidence limits. We assessed clinical and laboratory measurements and medication use from baseline through 10 years using generalized linear regression and generalized estimating equations. We performed stratified analyses according to the intention-to-treat principle. These stratified models are post hoc, not prespecified in the original Look AHEAD protocol. After identifying evidence for differential intervention effects, we conducted post hoc sensitivity analyses restricted to individuals from three of the four diabetes subgroups to determine the effect of the ILI on risk for cardiovascular events among these individuals combined. We used SAS, version 9.4 (SAS Institute, Cary, NC) for analysis.

We characterized four diabetes subgroups. Qualitative hallmarks of the four subgroups included older age at diabetes onset (hereafter, older onset), poor glucose control, severe obesity, and younger age at diabetes onset (hereafter, younger onset). The four diabetes subgroups constituted 42% (older onset, n = 2,178), 14% (poor glucose control, n = 716), 24% (severe obesity, n = 1,222), and 20% (younger onset, n = 1,029) of the total trial cohort. Within subgroup, equal allocation of intervention arm was maintained (Table 1). The distributions of sex, race/ethnicity, and educational attainment differed across subgroup, respectively. Women were more likely to be in the younger-onset subgroup and less likely to be in the older-onset subgroup than expected. Use of diabetes medications and insulin was lowest for the older-onset subgroup and highest in the poor-glucose-control subgroup. The poor-glucose-control subgroup had the most atherosclerotic lipid profile. Prior history of CVD within subgroup ranged in proportion from 12% to 15%. The proportion of individuals with imputed cluster characteristics did not differ substantially across subgroup (0.6–1.0%).

There were 821 events for the primary cardiovascular outcome over 43,851 person-years of total follow-up (median, 9.4 years). Among the subgroups, the unadjusted incidence rate for the primary outcome per 1,000 person-years was 18.9 for older onset, 23.6 for poor glucose control, 18.2 for severe obesity, and 15.7 for younger onset (Table 2). The P value for the interaction term between intervention and diabetes subgroup was 0.0014, and we continued with stratified models. Among participants in the poor-glucose-control subgroup, random assignment to the ILI was associated with 85% higher HRs for the primary outcome (HR 1.85; 95% CI 1.32, 2.61) (Fig. 1). Marked separation of the cumulative hazard curve between the ILI and DSE groups among the poor-glucose-control subgroup occurred at year 3 (Fig. 2). Among the older-onset, severe-obesity, and younger-onset subgroups, random assignment to the ILI was associated with HR ≤0.90, but all 95% CIs included the null value. The P value was <0.05 for the interaction term between intervention and diabetes subgroup for the second (P = 0.04) and third (P = 0.046) composite secondary outcomes, but not for the first composite secondary outcome (P = 0.11). Because there was evidence for an interaction between the ILI and diabetes subgroups for three of the four prespecified trial outcomes and the overlap of individual CVD outcomes in the composite outcomes, we stratified all analyses to better understand the differential risk by subgroup. We observed a similar pattern for the effect of ILI on each of the secondary outcome as was reported for the primary outcome.

We assessed the effect of the ILI on the individual outcomes of death, myocardial infarction, hospitalized angina, stroke, heart failure, CABG, carotid endarterectomy, and peripheral vascular disease within diabetes subgroups (Supplementary Table 1). The incidence for each of the individual outcomes was higher for the DSE group than in the ILI group within the older-onset, severe-obesity, and younger-onset subgroups, with the exception of carotid endarterectomy and peripheral vascular disease, which were rare events. For the poor-glucose-control subgroup, the incidence of myocardial infarction, hospitalized angina, stroke, CABG, and carotid endarterectomy was higher in the ILI group than in the DSE group, and the absolute risk difference was greatest for hospitalized angina. The corresponding HR comparing ILI to DSE for each of these individual outcomes was >1 among the poor-glucose-control subgroup (Supplementary Table 2). Although the 95% CIs were wide, the direction of effect for the ILI on the HR for death was consistently toward benefit among each diabetes subgroup.

Among all diabetes subgroups, participants randomly assigned to the ILI had greater improvements in risk factors except LDL cholesterol level, compared with DSE (Supplementary Figs. 1–7). Improvements for each of these measurements for the ILI arm tended to be greatest at year 1 for each subgroup. For weight and waist circumference, between-arm differences persisted until year 10 for each subgroup with the exception of the poor-glucose-control subgroup (year 7). Between-arm differences in HbA_1c were sustained longest for the younger-onset subgroup and least sustained for the poor-glucose-control subgroup, where mean HbA_1c of the two arms converged by year 3. Between-arm differences in SBP were present beyond year 1 for the older- and younger-onset subgroups but not the poor-glucose-control and severe-obesity subgroups. During follow-up, the prevalence of use of diabetes medications, insulin, cholesterol medications, and hypertension medications was lower for individuals randomly assigned to the ILI compared with those in the DSE group among all diabetes subgroups. We did not observe differences in any association between the ILI and medication use by subgroup.

After observing increased risk for each composite cardiovascular outcome only among the poor-glucose-control subgroup, we performed a sensitivity analysis that excluded this subgroup and pooled the three remaining subgroups (sample n = 4,429). In this subsample of 86% of the cohort, the ILI was associated with 15% reduced risk for the primary cardiovascular outcome, as well as reduced risk for secondary outcome 2 and myocardial infarction (Supplementary Fig. 8). The direction of effect for the ILI was consistently toward benefit for all outcomes except peripheral vascular disease, though the 95% CIs included the null for most.

In this secondary analysis of data from the Look AHEAD trial of an ILI among overweight and obese individuals with T2D, we characterized diabetes subgroups and assessed for differential treatment effects due to a lifestyle intervention for weight loss. We identified four unique diabetes subgroups at the time of randomization that were based on the composition of the cohort and characterized by older age at diabetes onset, poor glucose control, severe obesity, and younger age at diabetes onset. We determined that the lifestyle intervention had differential effects on cardiovascular outcomes depending on a patient’s diabetes subgroup. Among only individuals in the poor-glucose-control diabetes subgroup, which composed 14% of the cohort, the lifestyle intervention was associated with an 85% greater risk for the primary cardiovascular outcome and greater risk for each of the secondary cardiovascular outcomes.

These findings are informative for implementing ILIs for weight loss among overweight and obese individuals with T2D for multiple reasons. First, although the majority of individuals with T2D are overweight, there are disease characteristics that distinguish unique disease subgroups. Second, these findings provide context to the original Look AHEAD findings of a null intervention effect for prevention of major fatal and nonfatal cardiovascular events (1). The original interpretation requires nuance (ie, the effect of the ILI on cardiovascular outcomes differed by disease subgroups). Among individuals in the poor-glucose-control subgroup, the ILI was consistently associated with higher risk for CVD. Among individuals from the remaining three diabetes subgroups combined (ie, older onset, severe obesity, and younger onset), the lifestyle intervention was associated with lower risk for CVD compared with the control arm. These three subgroups constituted the overwhelming majority of individuals in the Look AHEAD cohort. Although the precision of the strength of the ILI effect was modest for these subgroups individually, the magnitude and direction of this effect (reduced risk) were consistent for the primary and secondary outcomes. When analyses were restricted to individuals in the older-onset, severe-obesity, and younger-onset diabetes subgroups, the ILI was associated with lower risk for CVD and estimates were more precise. A third informative point is that the heightened risk for the cardiovascular outcomes among the poor-glucose-control group was not evident until 3 years after randomization and was not observed for the individual incident outcome of death.

An increasing number of studies have characterized diabetes subtypes, some with populations that include individuals with both type 1 and type 2 diabetes (6–11). In the present study, and similar to others, we observed subgroups hallmarked by high HbA_1c, severe obesity, and older age at onset. Others have been able to classify subgroups characterized by impairments in β-cell function and insulin resistance (6–8,10). We were unable to examine these characteristics, because treatment with insulin, common in Look AHEAD participants, would hinder the evaluation of these factors. Dennis et al. (11) observed differential treatment response from glucose-lowering therapies by diabetes subgroup on HbA_1c reduction but not for clinical outcomes of CVD hospitalization or death.

Our work adds to two recent data-driven assessments of heterogeneous treatment effects of the Look AHEAD intervention on the primary composite outcome. Using decision-tree modeling to partition the sample into six subgroups, Baum et al. (12) observed that the ILI was beneficial for two subgroups of individuals: those with HbA_1c ≥6.8% and those with HbA_1c <6.8% and a high self-reported health score. Individuals with HbA_1c <6.8% and a low self-reported health score actually experienced worse outcomes. The interpretation of these findings to the broader population of overweight individuals with T2D are somewhat unclear because investigators excluded 4.7% of participants before subgroup derivation and then restricted estimation of treatment effects to two of the six subgroups (40% of the analytic sample). de Vries et al. (13) estimated 10-year, individual predicted intervention effects on major cardiovascular events using penalized Cox regression. Prediction models for CVD incidence included covariate adjustment for demographics, health behaviors, clinical risk factors, microvascular complications, and all possible treatment-by-covariate interaction terms. Because this predicted treatment effect model was derived with Look AHEAD data, there is the possibility of over-fitting and type 1 false discoveries. Investigators observed a dose response for CVD prevention across quartiles of predicted treatment effect where the ILI was associated with CVD prevention among those with low predicted treatment effect (quartile 1) and greater CVD rates for the highest quartile of predicted treatment effect (quartile 4). Individuals in quartile 4 had the highest HbA_1c, similar to the poor-glucose-control diabetes subgroup characterized in our work. However, the composition of these two groups differs notably by prior history of CVD, which was lower in the poor-glucose-control subgroup in our study (13% vs. 37%). Espeland et al. (15), recently reported potential for heterogeneity of intervention effects on a multimorbidity composite outcome including eight age-related chronic diseases; Look AHEAD participants with fewer comorbid conditions received the greatest benefit.

The mechanism by which the ILI was associated with a higher risk for cardiovascular outcomes only among the poor-glucose-control diabetes subgroup is not clear. Participants in this subgroup had a mean HbA_1c >9% at baseline, far exceeding standard targets for glycemic control. Although the ILI arm among the poor-glucose-control subgroup had lower HbA_1c at years 1 and 2 than the DSE group, by year 3, the mean HbA_1c was 8.2% for both arms in this subgroup and was substantively higher than all other diabetes subgroups. The convergence of HbA_1c by year 3 among the poor-glucose-control subgroup coincided with an increase in the cumulative hazard curve for the ILI compared with DSE. Although intensive glucose lowering was associated with increased mortality overall in the Action to Control Cardiovascular Risk in Diabetes trial, the intervention was associated with CVD risk reduction for those with baseline HbA_1c ≤8.0% but not >8.0% (23). This collective body of research highlights a clinical gap of ineffective interventions among individuals with high HbA_1c.

There are several limitations of our work that merit discussion. First, as for any behavioral intervention, Look AHEAD participants were not masked to their intervention arm and may have sought off-study weight loss and treatment strategies, a particular concern for the control group. Second, participants were overweight or obese at baseline, motivated to lose weight, and capable of completing a maximal fitness test. Our results may only generalize to individuals with similar characteristics. Compared with national estimates of the time, the Look AHEAD population had a greater burden of obesity and hypertension but more favorable lipid and glucose profiles at baseline than the surveyed US population with diabetes (24). Third, we defined diabetes subgroups at a single point in time among prevalent diabetes cases, using characteristics that may be influenced by disease progression, not disease etiology. Whether an individual would be categorized to a different subgroup on the basis of their clustering profile at diabetes onset is not clear and we do not assert that these subgroups are absolute. We assigned individuals to the single most probable diabetes subgroup on the basis of their entire clustering profile, but participants could share defining characteristics of other subgroups. Prior work characterizing diabetes subgroups has supported similar and different numbers of subgroups, though some studies included individuals considered to have traditional type 1 diabetes who were excluded from Look AHEAD (5–9). We assessed for differential intervention effects among subgroups that were not considered during the Look AHEAD trial design. Randomization was balanced within subgroup and we maintained the intention-to-treat analysis approach. However, our analyses are exploratory and post hoc and require confirmation elsewhere.

Our findings help disentangle heterogeneity in the impact of the lifestyle intervention from the Look AHEAD trial. Specifically, we observed increased risk of CVD from the intervention in a subgroup of the trial participants with poor glucose control at baseline. The overwhelming majority of participants did not experience an increased rate of CVD from the intervention and actually received benefit for multiple cardiovascular risk factors. This work supports consideration for lifestyle interventions for weight loss in diabetes that are targeted to subgroups of individuals to reduce potential harm and optimize benefit. These results provide support to investigate whether our findings apply to other diabetes complications and to assess what interventions would be beneficial to those individuals where the Look AHEAD intervention caused CVD-associated morbidity and death.

Acknowledgments. The authors thank the other investigators, the staff, and the participants of the Look AHEAD trial for their valuable contributions. A full list of participating investigators and institutions can be found at https://www.lookaheadtrial.org/. This trial is registered at ClinicalTrials.Gov (identifier: NCT00017953; https://clinicaltrials.gov/ct2/show/NCT00017953. The data used for analysis during the present study were housed at the data coordinating center and are not available for public distribution. However, all data used for this analysis were supplied by Look AHEAD investigators to the National Institute of Diabetes and Digestive and Kidney Diseases Central Repository and are publicly available at https://repository.niddk.nih.gov/studies/look-ahead/. Some of the information contained herein was derived from data provided by the Bureau of Vital Statistics, New York City Department of Health and Mental Hygiene.

Funding. This study used data from the Look AHEAD trial intervention phase, which was funded by the National Institutes of Health (NIH) through the following cooperative agreements with the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK): DK57136, DK57149, DK56990, DK57177, DK57171, DK57151, DK57182, DK57131, DK57002, DK57078, DK57154, DK57178, DK57219, DK57008, DK57135, and DK56992. Additional funding was provided by the National Heart, Lung, and Blood Institute, National Institute of Nursing Research, National Center on Minority Health and Health Disparities, NIH Office of Research on Women’s Health, and the Centers for Disease Control and Prevention. This research was supported in part by the Intramural Research Program of the NIDDK. The Indian Health Service provided personnel, medical oversight, and use of facilities. Additional support was received from The Johns Hopkins Medical Institutions Bayview General Clinical Research Center (M01RR02719); the Massachusetts General Hospital Mallinckrodt General Clinical Research Center and the Massachusetts Institute of Technology General Clinical Research Center (M01RR01066); the Harvard Clinical and Translational Science Center (RR025758-04); the University of Colorado Health Sciences Center General Clinical Research Center (M01RR00051) and Clinical Nutrition Research Unit (P30 DK48520); the University of Tennessee at Memphis General Clinical Research Center (M01RR0021140); the University of Pittsburgh General Clinical Research Center (M01RR000056); the Clinical Translational Research Center funded by the Clinical and Translational Science Award (UL1 RR 024153) and NIH grant (DK 046204); the VA Puget Sound Health Care System Medical Research Service, Department of Veterans Affairs; and the Frederic C. Bartter General Clinical Research Center (M01RR01346). Additional funding for individual investigators for this current work was provided by the NIDDK (grant DK110341, E.V., principal investigator).

The views expressed in this manuscript are those of the authors and do not necessarily represent the views of the National Institute of Diabetes and Digestive and Kidney Diseases, the National Institutes of Health, the U.S. Department of Health and Human Services, or the Indian Health Service.

Duality of Interest. The following organizations have committed to make major contributions to Look AHEAD: FedEx Corporation, Health Management Resources, LifeScan, Inc., a Johnson & Johnson Company, OPTIFAST of Nestle HealthCare Nutrition, Inc., Hoffmann-La Roche Inc., Abbott Nutrition, and Slim-Fast Brand of Unilever North America. No other potential conflicts of interest relevant to this article were reported.

Author Contributions. All authors made substantial intellectual contributions participating in creating and designing the study, analyzing and interpreting the data, and reviewing this manuscript. All authors have read and approved the final report for publication. M.P.B. is the guarantor of this work and, as such, had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.