Prediction of Coronary Heart Disease in Middle-Aged Adults With Diabetes

The ARIC study is a prospective study of cardiovascular disease in a cohort of 15,792 people, aged 45–64 years in 1987–1989, sampled from four U.S. communities (7,8). The cohort continues to be followed for morbidity and mortality, and this report includes follow-up through 1998, for a median of 10.2 years. A recent study has described CHD risk prediction in the entire cohort using basic and novel risk factors (9). For this study, participants who were not black or white (n = 48) and those who were black in two centers (n = 55) were excluded from analysis because of small numbers. Also excluded were participants who had preexisting CHD at baseline (n = 763), were missing data on baseline preexisting CHD (n = 339), or were missing data on one of the basic risk factors or BMI (n = 533), leaving 14,054 individuals available for analysis, 1,500 of whom had diabetes.

Details of ARIC measurements have been reported elsewhere (7–9), including the sports index (10) and Keys scores from diet (11). Residual forced expiratory volume in 1 s (FEV₁) was calculated as the difference between the observed FEV₁ and FEV₁ predicted from age, height, and sex. A 12-lead electrocardiogram (ECG) was used to define left ventricular hypertrophy (LVH) using the Cornell score (12). The ankle-brachial index (ABI) was measured, and peripheral arterial disease (PAD) was defined as ABI ≤0.90 for men and ≤0.85 for women. ARIC’s carotid ultrasound measurements used a scanning protocol common to the four field centers (13) and standardized central reading of scans (14). Analyses are based on mean intima-media thickness (IMT) of the far wall for 1-cm lengths of the carotid bifurcation and the internal and common carotid, right and left. The means at the six sites were combined in an unweighted average to produce an overall mean IMT. High IMT was defined as mean IMT ≥1.0 mm.

Prevalent diabetes was defined according to American Diabetes Association (ADA) criteria (15) as a fasting glucose level ≥126 mg/dl, nonfasting level ≥200 mg/dl, self-reported physician diagnosis of diabetes, or pharmacological treatment for diabetes. Among people who knew they had diabetes, 96% reported onset after age 30 years, so most ARIC diabetic participants presumably have type 2 diabetes.

CHD incidence in ARIC was ascertained as previously described (8,16). A CHD event was defined as a validated definite or probable hospitalized myocardial infarction, a definite CHD death, an unrecognized myocardial infarction defined by ARIC ECG readings, or coronary revascularization.

Descriptive statistics were computed by the level of each basic risk factor, including age-adjusted CHD event rates calculated from Poisson regression. From the coefficients of a proportional hazards model of time to incident CHD, a CHD risk score was calculated for each person by multiplying each model coefficient times the person’s level for the risk variable associated with that coefficient, then summing these products. The CHD risk score was created two ways: 1) by sex-specific regression modeling within the whole population, including diabetes status in the risk score (similar to the Framingham Study) (1), and 2) by performing sex-specific regression modeling separately for those with and without diabetes (similar to the U.K. Prospective Diabetes Study) (3). Thus, for men and women, the former predicted CHD risk in the full ARIC cohort, giving diabetic participants the higher predicted CHD risk because of their diabetes, and the latter predicted CHD risk based on the cohort with diabetes separate from the cohort without diabetes.

Our measure of individual risk predictivity of a model was the area under the receiver operating characteristic (ROC) curve (AUC) (17), which ranges between 0.5 and 1 and represents the probability that a person who had an incident event by a given follow-up time had a higher risk score than a person who did not have an event by that time. The area is a function of time; we used 10 years throughout and used Kaplan-Meier-like methods to calculate the relevant probabilities of event by that time in the face of censoring. Because many potential risk factors were studied and some parsimony was desirable, we used a 0.0025 change in AUC, in any subgroup, to include (Tables 2 and 3) or exclude (Table 4) a risk factor from the prediction model. The hypothesis “risk score A yields a higher AUC than risk score B” was tested via bootstrapping (17,18). The AUC for the basic model varied slightly among Tables 2–4 because of different sets of missing data.

We present graphs of the predicted probability of a CHD event within the first 10 years of follow-up by decile of risk score. The predicted probability of incident CHD within 10 years of follow-up was calculated by decile as the mean predicted probability at 10 years over all persons in the decile. Improved prediction was indicated by moving more of the predicted events out of the lower deciles of risk score into the upper deciles. All models showed good agreement between observed and expected number of events by decile of risk score, using the Hosmer-Lemeshow statistic (19).

Among ARIC participants aged 45–64 years with diabetes at baseline, 47% reported taking glucose-lowering medication (21% insulin, 26% other), 44% were untreated but met glucose criteria for diabetes, and 19% reported physician-diagnosed diabetes, but were normoglycemic on the day of examination. Among the diabetic participants, 55% were white and 45% were black.

Incidence rates of CHD were ∼4 times higher in diabetic than nondiabetic women and 2.5 times higher in diabetic than nondiabetic men (Table 1). However, absolute differences in CHD rates between diabetic and nondiabetic participants were greater for men than for women. CHD rates were greater for diabetic than nondiabetic participants within each level of basic risk factors and subclinical disease markers (Table 1).

Our first question noted above concerning how well basic risk factors identify diabetic participants who will or will not develop CHD was addressed using the AUC of ROC curves. We first predicted CHD using a “basic” risk score based on all ARIC participants and included a term to indicate diabetes or not. For comparison, we then predicted CHD in diabetes-specific models using the same basic risk factors. For participants with diabetes, particularly men, the specific model (AUC = 0.72 for women and 0.67 for men) was better than the combined model (AUC = 0.71 for women and 0.65 for men) at predicting CHD. Therefore, the remaining results focus on diabetes-specific models.

As shown in Table 2 (row 7), using diabetes status-specific “basic” risk scores alone, predictivity was lower in diabetic than in nondiabetic women (AUC = 0.72 vs. 0.79). In contrast, predictivity of basic risk factors was similar in diabetic and nondiabetic men (AUC = 0.67 vs. 0.69). Table 2 (rows 3–6) also indicates that the basic risk factors whose contribution to predictivity would most be missed, if removed, were smoking, total cholesterol, and blood pressure in diabetic women, and total and HDL cholesterol in diabetic men. When we used more restrictive definitions of diabetes (i.e., only drug-treated diabetes or only physician-diagnosed or -treated diabetes), the AUC values for the basic model were slightly lower for diabetic women (AUC = 0.70 for both restrictive definitions) and slightly higher for diabetic men (AUC = 0.75 and 0.72, respectively).

The other information in Tables 2–4 addresses our second study question of whether nontraditional risk factors or subclinical markers of disease improve prediction of CHD in diabetic individuals. As is shown in Table 2, based on diabetes status-specific models, single risk factors added one at a time to the “basic” model increased the AUC only modestly. For diabetic women, the addition of BMI, sport index, serum creatinine, white blood cell count (WBC), factor VIII, or the von Willebrand factor raised the AUC to ≥0.73. For diabetic men, the addition of waist-to-hip ratio, Keys score, factor VIII, or von Willebrand factor raised the AUC to >0.68. As shown in the bottom row, consideration of multiple risk factors raised these AUCs further to 0.77 (a 7% increase) in diabetic women and 0.74 (a 10% increase) in diabetic men. This gain in AUC from the basic model to the fullest model was larger in diabetic than in nondiabetic participants.

In diabetic and nondiabetic participants, the gain in AUC from adding subclinical markers one at a time to the basic risk factors was most substantial for carotid IMT (Table 3). The AUC increased to 0.72 (∼2%) for diabetic women and 0.70 (∼3%) for diabetic men using basic risk factors and dichotomous indicators of carotid IMT, ECG LVH, and PAD.

As shown in Table 4, using all relevant basic and nontraditional risk factors and subclinical markers, the AUC increased by ∼10% in diabetic participants, a far larger increase than the increases of ≤5% for nondiabetic participants. Because the maximum AUC is 1.0, an AUC increase of the size observed also represents ∼25% of the available AUC not accounted for using the basic risk factor model. Stated differently, the additional variables increased by 0.067–0.081 the probability that diabetic participants with incident CHD had a higher risk score than those without CHD (Table 4). In each subgroup, the final model provided significantly better prediction of CHD than the basic model (P < 0.05). For diabetic participants, however, it would be possible to drop lipoprotein(a) and fibrinogen without affecting the AUC of the final model in Table 4, thus leaving BMI, waist-to-hip ratio, sports activity, Keys score, serum creatinine, serum albumin, factor VIII, WBC, ECG LVH, and carotid IMT as the factors responsible for the increase in AUC beyond the basic model. The prediction models are presented in the 1appendix, and an interactive user-friendly risk calculator is found at www.aricnews.net, in which one can enter a diabetic patient’s risk factor levels and receive a predicted 10-year risk of CHD.

Some treated diabetic participants did not fast 8 h before their blood was drawn. We therefore had to run supplemental sex-specific models on fasted participants only to consider whether baseline concentrations of serum glucose, plasma triglycerides, or treatment status (insulin, oral hypoglycemic agents, or no pharmacological therapy) added further to our CHD prediction models. Starting with the final models from Table 4, the AUC in diabetic women increased by 0.005 after adding triglycerides, but increased inconsequentially by adding treatment status or fasting glucose to the model. In diabetic men, triglycerides did not add substantially to the AUC, but treatment status added 0.017 and fasting glucose added 0.010. The multivariately adjusted relative risk of CHD for diabetic men using insulin was ∼2.9 and for oral agents was ∼1.6, compared with no pharmacological therapy.

Figure 1 addresses our third study question about the probability of CHD at various levels of diabetes-specific risk scores. As expected, CHD risks were higher for participants with versus those without diabetes. For diabetic men and women, rates of CHD rose ∼10-fold from the lowest to highest deciles of the basic risk score (Fig. 1, dotted line). About half (53%) of diabetic men and 23% of diabetic women had a predicted 10-year CHD probability ≥20%, and ∼86% of diabetic men and 61% of diabetic women had a predicted 10-year CHD probability ≥10%. Based on the basic risk score, the average 10-year CHD probability among diabetic participants was 26% for white men, 15% for black men, 17% for white women, and 13% for black women.

The addition of nontraditional risk factors and subclinical markers to risk scores increased the predictivity of CHD, especially among diabetic participants (Fig. 1, solid line); that is, the CHD probability became much higher in the highest deciles of risk, based on the full compared with the basic model, and correspondingly became lower for the lowest deciles. Diabetic men and women in the highest decile of the full risk score had a 10-year probability of CHD >50%.

People with diabetes are at high risk of CHD, so much so that the NCEP has recommended that diabetic patients not receive specific CHD risk assessment, but have their plasma lipids managed as if they had CHD (4). The NCEP appears to have based its recommendation on Finnish data suggesting that diabetic patients with no history of myocardial infarction have a risk of a future myocardial infarction similar to that of nondiabetic patients with CHD (20), a finding not confirmed in Scotland (21) or in U.S. male physicians (22). Our data verified that, compared with nondiabetic individuals, most people with diabetes have elevated risk of incident CHD (at least 1% per year). However, ∼14% of diabetic men and 39% of diabetic women in ARIC had a predicted probability of CHD <1% per year when stratified by basic CHD risk score. This variation in risk may in part reflect the fact that many ARIC diabetic participants were defined by epidemiological criteria, which require only one blood glucose abnormality, and 36% were newly recognized by fasting glucose criteria alone. Yet, ARIC CHD rates are comparable to those in the U.K. Prospective Diabetes Study, where the CHD risk was ∼1.5% per year (3). Taken together, these data suggest that there may still be a role for CHD risk stratification of diabetic patients in preventive medical practice. Patients with lowest CHD risk might need less aggressive intervention than those at high risk.

Our study results corroborated the findings of previous large studies (23–26) for an important role of the basic risk factors in determining CHD incidence in diabetic individuals. Our data suggested, not surprisingly, that the basic risk factors predicted CHD incidence somewhat better using a prediction model like that used in the U.K. Prospective Diabetes Study (3), derived from diabetic participants only, than using that from the Framingham study (1), a model derived from the whole ARIC sample with diabetes (yes/no) as a term in the prediction model. With increased availability of web-based or personal digital assistant calculators, risk formulas for particular patient subgroups become increasingly feasible. Indeed, the “basic” equation in the 1appendix (Table A) has been adapted so clinicians can predict CHD risk in patients with diabetes comparable with those in ARIC (see ARIC website, www.aricnews.net). Although the basic risk factors did not predict CHD quite as well in diabetic participants as in nondiabetic participants, the absolute differences in CHD incidence across basic risk factors was greater for diabetic than nondiabetic participants, emphasizing the great potential for CHD prevention in diabetic individuals. Further rationale for control of the basic CHD risk factors in diabetic patients has been reviewed elsewhere (27).

The American Heart Association’s Prevention Conference VI concluded that data were scanty on whether nontraditional risk factors or subclinical disease markers could help further in CHD risk assessment in diabetic patients (6). That conclusion largely motivated us to explore additional risk markers in ARIC participants with diabetes, as we had in the whole ARIC cohort (9). The nontraditional risk factors or markers we examined individually contributed little to the AUC beyond the basic risk factors, but the addition of multiple markers increased the AUC by ∼10%. Notably, because the AUC for the basic risk factor model was lower for diabetic than nondiabetic participants, the increase in AUC by adding nontraditional risk markers was greater for participants with diabetes than those without diabetes. Whether these nontraditional measures would be useful and cost-effective in the care of diabetic patients seems unlikely. The anthropometric, blood, and ECG measurements required by the full risk equation (1appendix) could be easily performed in clinical practice, but the assessment of exercise, diet, and carotid IMT require more effort and cost.

Although our study was large, prospective, and population based, and carefully assessed CHD events, it had some shortcomings. First, as mentioned above, the diagnosis was an epidemiological, not clinical, one. An oral glucose tolerance test was not performed and diabetes was defined by ADA criteria (15). Second, we measured many candidate risk markers, but not HbA_1c, C-reactive protein, microalbuminuria, fibrinolytic factors, or coronary vascular calcium. These may have improved prediction further. Third, some of our survival models included a large number of variables relative to the number of events, especially for diabetic participants, potentially leading to imprecise estimates of the model coefficients. We also did not have enough events in diabetic participants to split the sample and perform a test analysis on half and validation on the other half. Future studies, therefore, would be useful for comparison. Finally, the final AUCs of 0.776 in diabetic women and 0.763 in diabetic men for the full model were still far enough from perfect prediction (AUC = 1.0) that better risk markers still could be sought; however, this lack of predictivity must partly reflect imprecision in characterizing risk factors.

Table A shows the coefficients of the risk factors for the CHD prediction equations from ARIC diabetic participants. The risk score XBETA for an individual is the sum of the products of the beta coefficients for a factor and the level of that factor for the individual. These risk scores can be converted by the following formula to probabilities (P) of onset of CHD within 10 years:

where S_m is 1 minus the probability of a CHD event within 10 years for a person with 0 levels of the categorical variables in the risk score and with the median values of the continuous risk factors, and XBETAmed is the risk score calculated for such a person. The median values, S_m, and XBETAmed are given in Table A.

To apply this risk score to a new population, even though the score may still properly rank individuals in the new population for CHD risk, one would need to calibrate P to the overall 10-year risk level in the new population. An approximate way to do this is to replace XBETAmed with a corresponding value calculated still at zero values of categorical variables but at the new population’s medians for the continuous variables, and to replace S_m with

where XBETAmed is for the new population, r is the average annual event rate for the new population, t is the years at which the risk probability is to be calculated, and mean (XBETA) is the mean risk score for the new population. The latter is estimated from summing the β-coefficients multiplied times the estimated means of the associated risk factors in the new population.

The Atherosclerosis Risk In Communities (ARIC) study was funded by contracts N01-HC-55015, N01-HC-55016, N01-HC-55018, N01-HC-55019, N01-HC-55020, N01-HC-55021, and N01-HC-55022 from the National Heart, Lung, and Blood Institute.