The Burden of Treatment Failure in Type 2 Diabetes

The study subjects were members of a not-for-profit, group-model, health maintenance organization, Kaiser Permanente Northwest (KPNW), located in Oregon and southwestern Washington state. Demographically and economically, KPNW’s 450,000 members resemble the area population (9): 47.0% male, 97.5% non-Hispanic, and 91.4% white. In 2002, 8.3% of the KPNW adult membership (>18 years of age) had diagnosed diabetes. For the time window used in this study, 1994–2002, HbA_1c averaged 7.6% across the entire KPNW registry. Nearly two-thirds (66.3%) of registrants had a mean HbA_1c of <8%, and 87.3% were <9% (10).

For this study, we selected all registrants who, between 1994 and 2002, initiated sulfonylurea monotherapy, metformin monotherapy, combination ther-apy, or were newly diagnosed and managed without drugs. To eliminate cases of primary (immediate) failure or intolerance, we excluded all registrants who did not maintain their new therapy for at least 12 months. Because we were studying only patients who experienced failure, we also eliminated all cases who did not ultimately change therapy during the study window. To demonstrate that a therapy had been newly begun, we required all subjects to have been a health plan member for at least 12 months before their first purchase or diagnosis. (Most subjects had been members for many prior years.) Some patients had time to initiate two or even three qualifying therapies during the study window and therefore appear in more than one analysis.

We searched electronic pharmacy records back through 1987 to ascertain drug-use history. These records are very complete: 95% of KPNW members with diabetes report getting all or most of their medicines for diabetes from KPNW pharmacies (data not shown).

All HbA_1c tests were performed by a single in-house laboratory using the Diamat assay, the standard method used in the Diabetes Control and Complications Trial (1). The normal range for this assay is 5.05 ± 0.5%. During the study window, the laboratory performed more than three ambulatory HbA_1c tests per average registrant per year, not counting test results obtained during hospital stays, during emergency room visits, or from portable glucometers.

To ensure complete histories of antihyperglycemic treatment, we included only individuals who had had their diabetes diagnosed while they were members of KPNW (80% of registrants). We assumed a member was diagnosed by KPNW if he or she had at least 1 full year of membership before diagnosis, without any of the following indications of diabetes: antihyperglycemic drug use, purchase of supplies for blood glucose testing, outpatient diagnostic notations of diabetes, inpatient discharge diagnoses of diabetes, participation in diabetes education, or a diagnostically elevated glucose or glycated hemoglobin level (using current ADA criteria [11]).

We defined total glycemic burden as the cumulative amount by which HbA_1c has exceeded a specified treatment goal or threshold, the area above a line. In this report, we calculated total glycemic burden as the sum of the differences between the patient’s actual or interpolated HbA_1c and 8.0% (or, alternatively, 7.0%), during each month that a patient was using a therapy under study and had an HbA_1c >8.0 or >7.0%. Figure 1 illustrates burden graphically for a hypothetical individual subject, using a treatment threshold of 8.0%, and shows the formula. To interpolate HbA_1c during months when no measurement was taken, we calculated the slope of the straight line connecting the two nearest measurements. We defined a second measure, avoidable glycemic burden, as the portion of total glycemic burden that occurs after the first HbA_1c test that revealed that a patient had exceeded goal or threshold (see Fig. 1).

A key uncertainty for physicians treating type 2 diabetes is whether future HbA_1c values will continue to deteriorate after a patient has crossed a threshold. To calculate the likelihood of further deterioration, we estimated the proportion of cases in which HbA_1c increased after each of 10 HbA_1c thresholds was first exceeded. In these analyses, we limited the study samples to subjects 1) who had at least one additional HbA_1c measurement between 2 and 12 months after crossing the threshold (and before initiating a successor therapy) and 2) who had at least one HbA_1c measurement after initiating the therapy but before crossing the threshold. These criteria exclude cases of primary failure as well as individuals who switched out of therapy before a follow-up measurement was taken.

Additionally, to define rules to assist clinicians in predicting when HbA_1c will continue to deteriorate, we estimated multivariate logistic regression models that predicted further deterioration. The dependent variable in these models was the log odds that HbA_1c would increase by >0.10 percentage points. Predictor variables were centered at their means and always included the immediately preceding HbA_1c value, age, sex, number of months on the therapy of interest, and, except in the case of nonpharmacologic therapy, an index of medication adherence. The adherence index was the ratio of the “days-supply” recorded for the prescription divided by mean observed days between medication refills.

To ascertain whether primary care physicians differed in their propensity to change therapy early in the secondary treatment failure process, we calculated, for each treatment category and for each KPNW primary care clinician who had at least five patients who failed therapy in the category, the proportion of their patients who were switched to a new or additional therapy before specified HbA_1c levels were reached. For each treatment category and HbA_1c threshold, we then performed statistical tests for homogeneity to test whether the observed pattern of physician behavior could have occurred by chance. Further analyses were planned if these tests disconfirmed the hypothesis that switching thresholds were random.

A total of 7,208 episodes of treatment initiation and eventual secondary failure qualified for inclusion in the study. Most of the episodes were in the diet-and-exercise (n = 2,319) and sulfonylurea monotherapy (n = 3,394) categories. Table 1 describes these episodes and the glycemic burden they imposed. The last HbA_1c value before starting therapy ranged from a high of 9.6% before sulfonylurea monotherapy (at the end of nonpharmacologic therapy) to 8.8% before metformin monotherapy (always, in these data, after a course of sulfonylurea monotherapy). The mean first HbA_1c value after treatment (or after diagnosis, in the case of nondrug therapy) was lowest for nonpharmacologic therapy (7.4%), followed by sulfonylurea monotherapy (7.6%), metformin monotherapy (8.2%), and, finally, sulfonylurea/metformin combination therapy (8.4%). Glucose control usually improved further with time, dropping ultimately to between 7.1% (sulfonylurea monotherapy) and 7.7% (metformin monotherapy and combination therapy). The last HbA_1c value before treatment abandonment or supplementation ranged from 8.6% in the nonpharmacologically treated group to 9.6% for combination therapy. The mean number of months that elapsed until a new or additional treatment was started ranged from 26.5 months (metformin monotherapy) to 35.1 months (sulfonylurea monotherapy).

Glycemic burden varied widely across treatments and reached high levels. When a threshold of HbA_1c 8.0% was used, total glycemic burden ranged from 8.6 HbA_1c-months for individuals on diet and exercise to 29.9 HbA_1c-months for individuals on combination therapy. Total burden for sulfonylurea and metformin monotherapy averaged 19.4 and 17.0 HbA_1c-months, respectively. Avoidable burden (at the 8.0% threshold) averaged 3.5 months less than total burden, a reduction that varied little across treatment groups.

Glycemic burdens were much greater when calculated from a base of 7.0%. Using this base, total glycemic burden (HbA_1c-months) averaged 22.5 for nonpharmacologic therapy, 41.5 for sulfonylurea monotherapy, 33.5 for metformin monotherapy, and 58.3 for combination therapy.

Most nonpharmacologically treated subjects (73.6%) switched to sulfonylurea therapy. (Although, in later years, metformin monotherapy became a common initial pharmacotherapy, few such patients had experienced treatment failure by the end of 2002.) Most sulfonylurea monotherapy episodes (91.3%) progressed to the substitution or, more typically, to the addition of metformin. Conversely, most metformin monotherapy episodes (91.6%) terminated with the addition of a sulfonylurea. Most combination therapy episodes ended with the initiation of insulin (87.7%).

Two-thirds (66.6%) of nonpharmacologically treated patients switched to an oral agent as soon as, or before, their HbA_1c first exceeded 8.0%. The corresponding proportions for sulfonylurea and metformin monotherapy were 35.3 and 44.6%. Only 18.6% of combination therapy subjects switched before exceeding this threshold.

Figure 2 displays, for 10 potential treatment-switching thresholds from HbA_1c 7.0 to 9.0%, the percentage of cases in which the next postthreshold HbA_1c result exceeded 8.0 and 7.0%. Out of necessity, these analyses include only patients who remained on their therapy after crossing a threshold. (These are the only patients whose follow-up test results could be observed.) Fig. 2 shows the likelihood that this failure to switch was potentially a poor decision because it was followed by further deterioration of HbA_1c. In general, as the switching threshold was increased, the likelihood that the next HbA_1c measurement would exceed 8.0% also increased. In all groups except the one for metformin monotherapy, at the ADA’s then-recommended treatment threshold of 8.0%, patients were as likely to experience an HbA_1c <8.0% at their next assay as they were to experience a >8.0% result. (For metformin monotherapy, the threshold of maximum possible uncertainty was slightly higher, 8.2%.) For patients on diet and exercise, sulfonylurea monotherapy, or metformin therapy, the probability that a patient’s next HbA_1c measurement would be <8.0% exceeded one-third, even when the switching threshold was set as high as 9.0%.

When we calculated whether the next HbA_1c would exceed 7.0% (the higher curves in Fig. 2), the likelihood of an elevated result tended to flatten out at higher switching thresholds, especially for nondrug therapy and sulfonylurea monotherapy. Even at the lowest switching threshold (7.0%), subsequent HbA_1c values stayed above 7.0% for a large majority of patients (59.9–69.1%). Consequently, we observed no 50/50 “threshold of maximum uncertainty” for any treatment or threshold level when a target value of 7.0% was specified. In addition, the probability that patients would return to an HbA_1c <7.0% once they experienced an HbA_1c ≥8.0% was small (0.126–0.254).

Table 2 shows the results of multivariate modeling of the log odds that the next HbA_1c measurement will increase by 0.10 percentage points for each treatment group. With the exception of male sex in the pharmacologically treated groups, all of the predictor coefficients were highly statistically significantly different from zero. In all models, a higher prior level of HbA_1c decreased the likelihood of a clinically significant increase in HbA_1c, after controlling for age, sex, months on therapy, and medication adherence. Older age decreased the odds of further deterioration of HbA_1c, as did evidence of better medication adherence. A longer duration in a therapy predicted higher odds of deterioration.

Despite these consistent, statistically significant results, the entropy statistics for these models are relatively low (6.7–9.2%), indicating that they predict only a small proportion of the total observed variation in change in HbA_1c. In homogeneity tests, we were unable to reject the null hypothesis that differences in clinicians’ propensity to initiate new treatment at various thresholds of HbA_1c were due to chance.

In type 2 diabetes, the successive failure of nonpharma-cologic therapy and oral antihyperglycemic agents eventually burdens patients with a heavy history of uncontrolled hyperglycemia. When this glycemic burden is defined as HbA_1c-years >8.0%, the total burden ranges from three-fourths of an HbA_1c-year (8.6 HbA_1c-months) in individuals failing diet and exercise to ∼2.5 HbA_1c-years (29.9 HbA_1c-months) for individuals failing combination therapy. When it is defined as HbA_1c-years >7.0%, the total burden is much higher, from 1.9 to 5.9 HbA_1c-years (22.5–58.3 HbA_1c-months). If, before starting insulin, a hypothetical patient were to progress from nonpharmacologic treatment through sulfonylurea or metformin monotherapy to combination oral agent therapy, as patients typically do in our study setting, he or she would accumulate nearly 5 HbA_1c-years of total burden >8.0% and about 10 HbA_1c-years of total burden >7.0%. The latter figure exceeds the mean reduction in glycemic burden (9.0 HbA_1c-years) achieved over 10 years by the U.K. Prospective Diabetes Study (3).

Our earlier retrospective study of KPNW members who switched to or added metformin after sulfonylurea therapy suggested that HbA_1c levels rise inexorably to levels well above 9.0% before clinicians respond (8). The present study confirms that HbA_1c levels reach at least 9.6% on average before combination therapy is attempted, but it also reveals a very “noisy” or jagged process of HbA_1c deterioration before these levels are reached. Perversely, the noise in this process is greatest at the very level (8.0%) at which (until 2004) clinicians have been asked to make new therapy decisions. In fact, when HbA_1c first crosses the formerly ADA-recommended treatment threshold of 8.0%, patients who are not immediately switched to new treatment are about as likely to experience a following result of <8.0% as they are to continue to deteriorate. This short-term unpredictability must make it almost impossible for clinicians to predict HbA_1c from test to test, when HbA_1c is allowed to approach 8.0%.

The present study has several important limitations. First, because a long time window is needed to observe the course of treatment failure, our results generally describe older treatment patterns, before thiazolidinediones, for example, were available for use and before metformin was used (in our setting) as a first-line agent. Also, patients were available for follow-up for various durations of time, depending on whether they initiated treatment earlier or later in the study window. Consequently, an average of 60% of subjects had not switched treatment by the end of the study window and were not, therefore, included in glycemic burden calculations. Our results therefore oversample patients who switched sooner rather than later and may therefore underestimate the total glycemic burden caused by failure. This is confirmed by the fact that the last HbA_1c value before the abandonment of nondrug therapy in episodes that ended during our study window averaged 9.6%, whereas the last HbA_1c value in episodes that both began and ended during our study window averaged 8.6%, a much lower number.

Another limitation is our inability to observe dosage increases as responses to loss of glycemic control. This may have contributed to our finding that, at higher thresholds of HbA_1c, subsequent HbA_1c tests were as or more likely to decrease as they were to increase. However, the impact of unobserved uptitration will have been limited by the fact that for the oral agents we studied, glyburide and metformin, uptitration yields only minimal additional glycemic control once recommended initial dosages are reached. Intensification of diet-and-exercise therapy could have had more significant effects, both during nonpharmacologic treatment and alongside oral agents. This, however, would not have alleviated much predictive uncertainty for clinicians because adherence to behavioral recommendations is itself hard to predict.

We found that terminal levels of HbA_1c increased substantially as patients progressed from nondrug therapy through combination therapy. The reasons for this cannot be identified in our data, but may include deterioration of β-cell function, probably exacerbated by lapses of metabolic control during treatment failure; increased insulin resistance, due to aging; increasingly sedentary lifestyle and weight-gain; psychological resistance to the use of insulin; loss of therapeutic motivation by patients or clinicians; and characteristics of the therapies themselves.

Our attempts to develop multivariate equations to aid clinicians in the prediction of future glycemic test results failed to explain most of the prospective variance in HbA_1c. A more practical aid would be the following simple rule: do not allow HbA_1c to get near 8.0% in the first place. An action threshold of 7.0% yields better predictive characteristics than a threshold of 8.0%. Between 60 and 70% of patients are likely to experience a continuing increase in HbA_1c after their results on this assay rise above 7.0%. If the target for glucose control is an HbA_1c <7.0%, our results strongly suggest that the recommended threshold for action should be 7.0% or lower. An even stronger signal would be provided by a treatment threshold of 6.0%, which has proved widely achievable in the test phase of the Action to Control Cardiovascular Risk in Diabetes (ACCORD) study (12).

This research was sponsored by GlaxoSmithKline and KPNW.