Predicting the Effect of Fenofibrate on Cardiovascular Risk for Individual Patients With Type 2 Diabetes

Patients with type 2 diabetes mellitus (T2DM) have a high risk of cardiovascular disease (CVD) (1). Although statins reduce cardiovascular morbidity and mortality significantly by reducing LDL cholesterol (LDL-C), the remaining residual risk underscores the clinical need for additional treatment options (2,3). Fibrates specifically target the dyslipidemia seen in T2DM by reducing plasma triglycerides (TG) by 30% and increasing HDL cholesterol (HDL-C) by 10% (4). The Fenofibrate Intervention and Event Lowering in Diabetes (FIELD) and the Action to Control Cardiovascular Risk in Diabetes (ACCORD) trials, both placebo-controlled randomized clinical trials, investigated the effect of fenofibrate in patients with T2DM and showed no overall reduction in CVD (5,6). However, although the benefit of fenofibrate was not statistically significant, it is conceivable that there are patients in whom fenofibrate does convey clinical benefit (3). Subgroup analyses are a method frequently used to address this issue. Based on subgroup analysis in FIELD and ACCORD, the American Diabetes Association guideline recommends statin/fibrate combination therapy in T2DM in men with TG >2.3 mmol/L and HDL-C <0.9 mmol/L (7).

Individualized treatment effect prediction considers both favorable and unfavorable risk factors simultaneously and takes relative effects presented by randomized trials one step further by expressing treatment effect in terms of absolute risk reduction (ARR) for individual patients (8–13).

The aim of the current study was to estimate individual treatment effect of fenofibrate on major cardiovascular events (MCVEs) for T2DM patients from the FIELD and ACCORD randomized clinical trials and the Second Manifestations of ARTerial disease (SMART) observational cohort study.

FIELD evaluated the effect of fenofibrate compared with placebo in 9,795 patients with T2DM between the age of 50 and 75 years who were not taking statin therapy at study entry. Detailed information about the study design has previously been published (3). The primary outcomes were coronary events (coronary disease mortality and nonfatal myocardial infarction), which occurred in 544 patients during a 5-year median follow-up. FIELD reported an 11% reduction in coronary events in patients using fenofibrate (hazard ratio [HR] 0.89 [95% CI 0.75–1.05]) (5). In total, 859 patients suffered an MCVE, defined as nonfatal myocardial infarction, nonfatal stroke, or cardiovascular mortality.

The ACCORD lipid trial consisted of 5,518 patients with T2DM and investigated the effect of fenofibrate in combination with atorvastatin compared with atorvastatin plus placebo (6,14). The primary outcome was MCVE, which occurred 601 times after a median follow-up of 4.7 years. The ACCORD trial reported an 8% reduction in MCVEs (i.e., nonfatal myocardial infarction, nonfatal stroke, and cardiovascular death) (HR 0.92 [95% CI 0.79–1.08]) in patients using fenofibrate on a background of statin therapy.

Although the primary end points used in FIELD and ACCORD differed, we used the MCVE composite end point for the analyses in the current study.

The SMART study is an ongoing longitudinal cohort study that started in 1996 and included 1,829 T2DM patients with and without CVD (15). In the SMART cohort, the composite end point MCVE was observed in 334 patients, with a median follow-up time of 5.8 years. In all three studies, written informed consent was obtained from all participants and the ethics boards of the institutions approved the studies.

In FIELD, 0.2% of data were missing for duration of T2DM and 0.3% for urinary albumin-to-creatinine ratio (UACR). In ACCORD, missing data ranged from 0.1% for current smoking to 4.3% in UACR. In SMART, missing data ranged from 0.05% for TG and HDL-C to 7% for UACR. Single imputation by bootstrapping and predictive mean matching was used (aregImpute in R, Hmisc package) to account for missing data in the predictors (16).

For development of the FIELD model for the prediction of 5-year ARR for MCVE by fenofibrate, Cox proportional hazards models for time to MCVE were used, with time to event in years. For prevention of overfitting, predictors were prespecified based on the presence in three or more CVD prediction models for patients with T2DM (17). The initial model included the following variables at baseline: fenofibrate treatment, age, sex, ethnicity, diabetes duration, current smoking, previous CVD, use of antihypertensive medication, HbA_1c, systolic blood pressure, non–HDL-C, HDL-C, TG, UACR, dyslipidemia, estimated glomerular filtration rate (eGFR) (MDRD formula), and interaction terms between these predictors and fenofibrate. Metabolic syndrome was not added as a separate categorical variable because its individual components were included as single predictors. In eight of the nine prognostic models that were used for predictor selection, a marker for body weight was initially included but fell out of the model owing to inadequate (additional) prognostic value. Therefore, we did not add BMI or waist circumference to our set of predictors. Fenofibrate use was forced in the model. Model selection was based on backward selection using the Wald χ² statistic for removal of variables, based on which ethnicity, TG, and dyslipidemia dropped out of the model. TG probably fell out of the model because their prognostic value is already captured in non–HDL-C, which is a stronger predictor for CVD. Continuous variables were checked for nonnormality by visual inspection of the Martingale residual distribution, which led to log transformation of UACR and HDL-C. The Cox proportional hazards assumption was tested and met for all predictors. The final model is shown in Supplementary Box 1.

Model validation was performed in the ACCORD lipid trial and the SMART observational cohort study. To ensure adequate risk stratification, we recalibrated the final model for baseline survival in ACCORD and SMART to take differences in baseline survival (e.g., by a difference in background lipid-lowering therapy) into account. Model performance was tested with calculation of the C-statistic for discrimination and with calibration plots for predicted versus observed MCVE risk (expressed as event-free survival). For comparison, the Action in Diabetes and Vascular Disease: Preterax and Diamicron MR Controlled Evaluation (ADVANCE) risk model (18) was validated in these three data sets (Supplementary Data).

For demonstration of the application of the prediction model in a large and heterogeneous group of patients that resembles clinical practice, patient data from FIELD, ACCORD, and SMART were combined (N = 17,142). Results were stratified for the presence of dyslipidemia, which was defined as TG >2.3 mmol/L and HDL-C <1.00 mmol/L for men and <1.30 mmol/L for women. Results stratified for TG ≤2.3 mmol/L and >2.3 mmol/L are shown in Supplementary Data. First, 5-year MCVE risk was calculated by setting treatment status to zero in the FIELD risk model for all patients (irrespective of their actual treatment status). Second, ARR was calculated as MCVE risk minus on-treatment risk. ARR (%) was also expressed as number needed to treat (NNT = 100/ARR). An example of how ARR is calculated can be found in Supplementary Box 2. Subgroup-specific HRs for the effect of fenofibrate on 5-year MCVE risk in patients with and without dyslipidemia were derived from the combined FIELD and ACCORD data using Cox proportional hazards models. These were HR 0.97 (95% CI 0.86–1.09) for patients without dyslipidemia and HR 0.77 (95% CI 0.64–0.94) for patients with dyslipidemia. On-treatment risk was calculated by multiplying 5-year MCVE risk with the subgroup-specific HR for patients with and without dyslipidemia. Third, patients were stratified according to treatment effect. The analyses were performed in R, version 3.2.2 (package ‘rms’; R Core Team, Vienna, Austria), and SAS, version 9.3. Fourth, a calculation sheet was made (Microsoft Office Excel 2007).

FIELD included 9,795 patients, ACCORD 5,518 patients, and SMART 1,829 patients with T2DM. The mean age was 62.2 ± 7.3 years in the total population, and 66% were men (Table 1). Of the FIELD participants, 22% had a history of CVD and the median duration of T2DM was 5 years (interquartile range [IQR] 2–10). In ACCORD, 37% of the patients had a history of CVD with a median duration of T2DM of 9 years (IQR 5–15). In SMART, 66% had a history of CVD and the median T2DM duration was 10 years (IQR 5–15 years).

The FIELD risk model for the 5-year treatment effect of fenofibrate on MCVE in patients with T2DM is provided in Supplementary Box 1. External validation showed moderate discrimination, with a C-statistic of 0.67 (95% CI 0.65–0.69) and 0.68 (95% 0.64–0.72) in ACCORD and SMART, respectively. Calibration between observed and predicted 5-year MCVE risk was well balanced in both studies (Fig. 1). The ADVANCE risk model was less well calibrated and had a lower C-statistic in ACCORD and SMART compared with the FIELD risk model (Supplementary Figs. 1–3).

In a pooled analysis of FIELD, ACCORD, and SMART, the median 5-year MCVE risk was 9.4% (IQR 5.4–16.2) and 6.7% (IQR 4.0–11.7) in patients with and patients without dyslipidemia, respectively (Fig. 2). There was a wide range in absolute treatment effect of fenofibrate, with a median ARR of 2.15% (IQR 1.24–3.69, NNT 47) in patients with dyslipidemia and 0.22% (IQR 0.13–0.38, NNT = 455) in patients without dyslipidemia (Fig. 2). Results were similar when stratified for TG ≤2.3 mmol/L and >2.3 mmol/L (Tables 1 and 2 and Supplementary Figs. 4 and 5). For calculation of individual ARR, a calculation sheet is provided in Supplementary Data.

Of patients with dyslipidemia, 54% had an ARR ≥2% (NNT ≤50). Patients with dyslipidemia and an ARR ≥2% (NNT ≤50) and had a median MCVE risk of 15.4% (IQR 11.7–22.5) and a median ARR of 3.52% (IQR 2.37–5.13) (Table 2). In patients with dyslipidemia and ARR <2%, the median MCVE risk was 5.2% (IQR 3.8–6.8) and the median ARR was 1.18% (0.86–1.55). Of patients with dyslipidemia, 16% (N = 634) had an ARR <1.0%. The median MCVE risk in these patients was 3.26% (IQR 2.63–3.90) and median ARR 0.74% (IQR 0.60–0.89, NNT = 135).

Of patients without dyslipidemia, 97% had ARR by fenofibrate of <1% (NNT >100). In these patients, the median MCVE risk was 6.6% (IQR 3.9–11.2) and ARR 0.21% (IQR 0.13–0.36) (Table 2). The median MCVE risk in patients without dyslipidemia with an ARR ≥1% (NNT ≤100) was 36.8% (IQR 33.6–42.9).

In this study, we used statistical modeling to estimate the individual treatment effect of fenofibrate for 17,142 patients with T2DM from the FIELD and ACCORD randomized trials and the SMART observational cohort study. We found a wide range of individual treatment effects. More than half (54%) of the patients with dyslipidemia had a substantial treatment effect (ARR >2%, NNT <50) and nearly all (97%) patients without dyslipidemia had a small treatment effect (<1%, NNT >100).

Individualized treatment prediction is a method to overcome disadvantages of subgroup analyses and facilitate the translation from randomized clinical trials to individual patients (8–13).

Subgroup analyses often overestimate treatment effect (19) and study only one patient characteristic at the time (17), thereby not taking into account that treatment effect usually depends on several, often correlated, patient characteristics. This can lead to confounded interpretations (20). Individualized treatment prediction uses multivariate models to estimate individual, absolute treatment effects, thereby contributing to the improvement of personalized medicine.

As mentioned before, patients with T2DM have a high residual cardiovascular risk, even when risk factors are adequately treated according to guidelines (2), and additional treatment options to reduce cardiovascular risk are needed (3). On the other hand, overtreatment is undesirable, especially in T2DM patients who usually use many medications including oral hyperglycemic drugs and/or insulin. Before adding additional drugs, such as fenofibrate, both physician and patient need to know what the expected effect of fenofibrate will be. The multivariable prediction model presented in this article can be used to translate group-level evidence to individual patients in clinical practice. Presenting an individual treatment effect can benefit communication with a patient because benefits can be weighed against expected harms such as costs and side effects. This will facilitate shared decision making and improve treatment adherence (21,22). Although the use of individualized treatment effect prediction in clinical practice might seem complicated, we provide an easy-to-use calculation sheet that only requires a physician to fill in the patient’s clinical information. This way physicians are able to practice evidence-based personalized medicine in daily practice.

External validation of the FIELD risk model showed good calibration and moderate discrimination in both the ACCORD randomized clinical trial and in “real-life patients” from the SMART observational cohort study after recalibration of baseline survival. This means that the FIELD risk model can be used to estimate 5-year MCVE risk for individual patients with T2DM. For individualized treatment effect prediction, it is especially important that the predicted risks are in agreement with the observed risks, and therefore calibration is the most important performance measure in this regard (10). One of the reasons that there was a difference in baseline survival might be the fact that FIELD patients were not on background therapy with statins. However, the well-balanced calibration of the FIELD model in ACCORD after baseline risk adjustment indicates that the difference in background statin therapy did not influence treatment effect predictions.

We found that, overall, patients with dyslipidemia had a larger treatment effect of fenofibrate than patients without dyslipidemia. This is due to a larger relative treatment effect, which is in line with two meta-analyses that found a higher relative benefit of fibrate in patients with dyslipidemia (23,24). Furthermore, patients with dyslipidemia generally have a higher 5-year MCVE risk than patients without dyslipidemia because both low HDL-C and high TG levels are independently related to increased CVD risk (25–27). That fenofibrate did not show an overall treatment effect in FIELD and ACCORD is probably because the majority of patients did not have dyslipidemia at baseline. Although we generally found a higher treatment effect in patients with dyslipidemia, we also found that some patients without dyslipidemia had a considerable fenofibrate treatment effect (ARR ≥1%, NNT <100). Furthermore, 16% of patients with dyslipidemia had a small treatment effect (ARR <1%, NNT >100). These results indicate that simply treating all patients with dyslipidemia and none without dyslipidemia with fenofibrate does not necessarily tailor the right treatment to the right patient. Therefore, treatment recommendations should go beyond group level and use individualized treatment effect prediction to guide clinical decision making.

Future projects to develop this method and add to the development of personalized medicine should be aimed at estimating the combined effect of different types of medication, e.g., glycemic control, statins, and blood pressure–lowering medication, on clinical end points because the combined effect of these medications might be different than each effect separately. For development of a model like this, trials that investigate the combined effect of these medications should be performed. Furthermore, the effect on microvascular T2DM end point such as nephropathy or retinopathy can be estimated, just like quality of life. A recent development is to predict gain in healthy life expectancy instead of risk because risk is in very large part determined by age (3). Another interesting future project would be to determine which lipid effect of fenofibrate (TG lowering, HDL increasing, or other) is most important for the observed treatment effect. Finally, as clinical practice changes over time, it would be valuable to validate and update the presented model regularly.

Some limitations with regard to this study should be considered. First, the prediction model presented here was developed using data from clinical trials, which might limit generalizability to patients with T2DM in clinical practice because clinical trial patients tend to have less comorbidity, better adherence, and a larger expected treatment benefit (28). However, validation of the FIELD risk model in the SMART observational cohort showed good risk calibration, indicating that generalization to real-life patients is probably fairly accurate. Furthermore, by using pooled data from FIELD, ACCORD, and SMART, we created a large and heterogeneous group of T2DM patients that resembles the wide variety of patients seen in clinical practice. Also, in FIELD there was a 23% statin drop-in rate; however, a calibration plot showed a good balance between predicted and observed risks in FIELD, indicating that the statin drop-in did not influence the precision of the predictions made at baseline (Supplementary Data). Second, the cutoffs for TG and HDL-C to define dyslipidemia have not been generally established. We used sex-specific HDL-C cutoffs (HDL-C <1.00 mmol/L for men and <1.30 mmol/L for women) as in the definition for metabolic syndrome (29) and as used in FIELD (5) and a TG cutoff of >2.3 mmol/L based on the definition for moderately increased TG of the 2012 Endocrine Society guideline (30), as well as the 2016 ESC/EAS (European Society of Cardiology and European Atherosclerosis Society) guidelines that recommend considering lipid-lowering treatment when TG levels are >2.3 mmol/L (31). The two meta-analyses that investigated the effect of fibrates on CVD also used the TG cutoff of >2.3 mmol/L, but one used an HDL-C cutoff of <0.9 mmol/L, and the other of <1.0 mmol/L, to define dyslipidemia (23,24). These differences in dyslipidemia definition might complicate the generalizability of our results. Third, predictions are made for a 5-year period, while patients and physicians might want information for a longer period or even for over the lifetime. The 5-year time period was chosen owing to the available follow-up data in the derivation and validation sets, but longer-term predictions would have been of added value.

In conclusion, in individual patients with T2DM there is a wide range in the individual treatment effect of fenofibrate on MCVE risk. We found a wide range of individual treatment effect of fenofibrate in patients with T2DM, and overall, patients with dyslipidemia had a larger fenofibrate treatment effect than patients without dyslipidemia. The method of individualized treatment effect prediction of fenofibrate on MCVE risk reduction in T2DM can be used to guide clinical decision making.

Acknowledgments. The authors gratefully acknowledge the contribution of Rachel O’Connell, statistician at the National Health and Medical Research Council Clinical Trials Centre.

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. C.K. researched data and wrote the manuscript. F.L.J.V. contributed to discussion and reviewed and edited the manuscript. J.W. reviewed and edited the manuscript. Y.v.d.G. contributed to the discussion and reviewed and edited the manuscript. H.N.G. reviewed and edited the manuscript. A.C.K. researched data, contributed to the discussion, and reviewed and edited the manuscript. F.L.J.V. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.