Previous trials have provided limited information on the variability of benefits and harms of treating to targets across the spectrum of the US diabetes patient population, making it difficult for clinicians to tailor their care to individual patients. We developed a simulation model using the best available evidence from clinical trials and found that diabetes patients with the highest CVD risk accounted for nearly all of the benefits of treating to targets while average-risk patients—nearly three-quarters of the population—received very little benefit. By accounting for treatment-related harms, we identified numerous examples in which intensifying treatment would be contraindicated on the basis of risk/benefit considerations, and many more instances in which the expected benefits are so small, that shared patient-clinician decision-making would appear to be the appropriate medical intervention. Of note, even for very high-risk patients with diabetes, increasing from standard to high doses of BP medications in pursuit of tight control goals was mainly effective for calcium channel blockers, calling into question the wisdom of titrating thiazides, ACE/ARBs or beta-blockers to high doses.
Results similar to our findings have been found frequently when heterogeneity of treatment effects are fully examined. Ioannidis noted over a decade ago that a small minority of high risk patients often account for the vast majority of outcomes in clinical trials.37
Many others have demonstrated how, under such circumstances, using average results from clinical trials can extend treatments to large numbers of patients who get little or no benefit, and sometimes net harm.8
The recent ACCORD trials highlight these risks. Greater use of polypharmacy in the intensive arm of the ACCORD-BP trial (SBP goal of less than 120 mm Hg) was associated with significantly higher rates of at least three types of serious adverse events, but provided no additional reduction in cardiovascular events compared to the non-intensive arm (SBP goal of less than 140 mm Hg). Although there are a number of possible explanations for the lack of benefit of intensive control, one likely explanation is that the majority of the benefit experienced by patients in both arms came from lowering patients’ BP from high levels to moderate levels, or in other words, from lowering blood pressure among the highest risk patients. While there was no overall benefit of combination therapy in the ACCORD-Lipid trial, a subgroup analysis suggested that there might be a benefit for subjects with the lowest levels of HDL and highest levels of triglycerides. Taken together, these results provide additional support for a more nuanced view of selecting risk factor target levels for patients with diabetes.
Our results highlight the implication of heterogeneity of treatment effects for current diabetes treatment guidelines, and that simply because the average CVD risk across all patients with diabetes is high, that does not mean that most people with diabetes have high risk. Most primary prevention guidelines are moving even more strongly to base recommendations on an individual patient's calculated CVD risk, and our results suggest that having diabetes should perhaps no longer be an exception to this general rule. Further, because LDL and blood pressure values are individually poor predictors of a person's overall CVD risk,38
many patients will receive little or no benefit when intensification is based solely on their current LDL and blood pressure levels, while at the same time, some high-risk patients might be undertreated. Disutility from side effects and high levels of polypharmacy have the potential to cause net harm in patients receiving little or no benefit from intensification, especially those who are already taking several medications— which is the majority of patients with diabetes today.39-41
For these patients, the next treatment is likely to have limited efficacy (due to the diminishing benefit of combination therapy), more side effects (because 3rd
, and 5th
-line agents and high doses tend to be less well tolerated), and a high polypharmacy burden. While the magnitude of the treatment harm might seem trivial, there are compelling reasons not to discount it. Greater use of polypharmacy is associated with a higher risk of drug interactions; more uncertain long-term safety risks (particularly when newer treatments are used); and significant cost, inconvenience, and side effect burdens that might engender higher rates of non-adherence. Given the large number of comorbidities often associated with diabetes, pursuing small marginal health gains through polypharmacy could have significant opportunity costs. We accounted for these factors using a small disutility that increased with the level of polypharmacy, but formally quantifying these effects is a challenge.
By attempting to simulate a complex clinical process, our model required a number of simplifications. There is no one standard treatment protocol for controlling LDL and blood pressure, and using a different set of treatments, additional treatments, or allowing therapeutic substitutions instead of simply additions or titrations, could impact our results. We restricted our focus to treatments likely to be considered standard in most practice settings and considered only therapies with known efficacy in lowering each risk factor. We did not consider fibrate therapy, for example, which is often prescribed to lower triglycerides and increase HDL, because it has limited efficacy in lowering LDL42
and has no clear effects on reducing CVD mortality.43
While we might have considered substituting therapies for patients having adverse events to lower the contribution of treatment harms, doing so would have resulted in lower rates of successful control and smaller health improvements.
Although we assessed all assumptions in sensitivity analyses, we were unable to find values for several model parameters in the clinical literature. We found no estimates for the relative blood pressure reduction observed with combination therapy involving three or more antihypertensives, so we extrapolated estimates from 2-drug combinations. The absence of these data is alarming given the high level of combination therapy used today. Lacking an estimate of the disutility for statin-induced myalgia, we assigned a value of 0.10; others have reported a higher disutility of 0.18 for myalgia following chemotherapy.44
Our myalgia incidence rates came from the only study we were able to find—a study of patients reporting muscular symptoms on high dose statins in an outpatient setting15
—and we assumed patients starting low dose statins had a rate half as large. We did not use myalgia rates reported in statin trials since these estimates are likely to significantly underestimate the true rates because of the use of runin phases for detecting intolerance45
or from the use of inclusion criteria requiring prior tolerance in these trials.16
Our results illustrate the complexities of decision making regarding control of CVD risk factors. The relative benefit of treatment depends on a patient's baseline risk factor level, underlying risk of adverse health outcomes, competing mortality risks, and current medications. It is possible that clinicians already incorporate the issues and principles addressed in our study to help make treatment decisions that are more in line with an individual patient's risks and benefits. Therefore, our model and assumptions may or may not represent how clinicians currently practice but rather models what would happen if current guidelines were followed rigidly by clinicians. Further, we are not proposing that clinicians should solely consider quantitative estimates of models such as ours when making individual patient decisions. However, having such estimates could greatly assist clinicians in helping their patients make personalized decisions, and given the complexity of the factors involved in estimating risks and benefits of lipid and blood pressure treatments, it seems likely that information systems that can assimilate this information and support evidence-based treatment recommendations will be needed for models like ours to be used in clinical practice. Decision support is even more important due to the speed with which the clinical literature evolves in this area.