Differential Treatment Response to Insulin Intensification Therapy: A Post Hoc Analysis of a Randomized Trial Comparing Premixed and Basal-Bolus Insulin Regimens

These results were a post hoc analysis of a 24-week, multicenter, open-label, randomized, parallel, active-controlled study comparing PPT with BBT [10].

This study included patients (a) with T2DM (b) who were 30–75 years of age, inclusive, at the time of screening, (c) who had been treated for at least 90 days with insulin glargine once daily (≥30 U/day) in combination with OAMs (sulfonylurea or meglitinide, metformin, or thiazolidinedione and, in the opinion of investigator, required further intensification of therapy, and (d) who had inadequate glycemic control [HbA1c ≥7.5% (58.5 mmol/mol) and ≤12.0% (107.7 mmol/mol)] at screening [10].

Major exclusion criteria were (a) patients who had severe hypoglycemia (more than one episode 6 months before study entry), (b) patients with excessive insulin resistance (receiving dose >2.0 U/kg at the time of randomization), (c) patients who were obese [body mass index (BMI) >45 kg/m²], (d) patients who had cardiac diseases with functional status Class III or IV as per the New York Heart Association [11], (e) patients who had renal disorders [serum creatinine ≥135 μmol/L (1.5 mg/dL) for men and ≥100 μmol/L (1.2 mg/dL) for women], or (f) patients who had liver disease (alanine aminotransferase at least three times the upper normal limit of reference range) [10].

Eligible patients, based on inclusion and exclusion criteria, were randomized to receive either PPT or BBT, and either to an aggressive dosing algorithm (based on plasma glucose levels and total daily insulin) or a conservative dosing algorithm (plasma glucose levels). Patients randomized to the PPT group received insulin lispro mix 50/50 (50% insulin lispro and 50% insulin lispro protamine suspension) three times daily before meals. If fasting blood glucose (FBG) targets were not achieved with insulin lispro mix 50/50, the evening preprandial dose could be changed to an injection of insulin lispro mix 75/25 (25% insulin lispro and 75% insulin lispro protamine suspension). The starting dose was based on the insulin glargine total dose at study entry divided into three equal doses of lispro mix 50/50 before meals. Patients randomized to the BBT group received insulin glargine once daily along with preprandial insulin lispro. The starting dose of this regimen was also based on the total glargine dose at study entry, with 50% administered as glargine and the other 50% given in three equal doses of lispro. Patients in both arms continued to receive OAMs; however, insulin secretagogues (sulfonylureas and meglitinides) were discontinued at randomization. The dose of insulin is decided with a goal of achieving preprandial blood glucose of less than 110 mg/dL (6.11 mmol/L) and a 2-h postprandial blood glucose of less than 140 mg/dL (7.78 mmol/L) with no hypoglycemia [blood glucose <72 mg/dL (4.0 mmol/L)]. Detailed study design, dosing schedule, dosing algorithm, efficacy, and safety results of this study were previously published [10].

All procedures followed were in accordance with the ethical standards of the responsible committee on human experimentation (institutional and national) and with the Helsinki Declaration of 1964, as revised in 2013. Informed consent was obtained from all patients for being included in the study.

Of 547 patients who entered the study, 374 were randomized by an interactive telephone system at a 1:1 ratio to the PPT and BBT groups, respectively [10]. Of these 374 patients, data for 182 patients (81 on PPT and 101 on BBT) who provided a baseline HbA1c and achieved a target 24-week HbA1c [<7.0% (53.0 mmol/mol); responders, also referred to as at-goal patients] were examined and explored. The data were analyzed by the SIDES algorithm [9], instead of specifying subgroup by anecdotal evidence, to explore a subgroup of patients (out of those 182 at-goal patients) who could show favorable responses to PPT.

The SIDES method, a novel recursive partitioning procedure, allows direct evaluation of the treatment effect in subgroups and is effective in evaluating modest-sized and large databases. This approach helps in searching for subgroups with a large differential effect relative to the study population. Also, this method can be applied as a part of prospectively defined tailored-therapy strategy and retrospectively throughout drug development across early- and late-phase trials combined to detect important predictive subgroups as early as possible [9]. The algorithm directly searches for large treatment effect simultaneously among many covariates. The traditional way of performing subgroup analyses involves prespecifying the subgroup of interests by specifying a set of variables and the corresponding cutoff points. The differences of treatment effects between comparator groups can then be examined by a parametric regression. Such an approach necessitates predefining subgroups of interests, which is often difficult when there are many covariates. Also, for any particular dataset, the subgroup suggested by our experience may not be the subgroup that had the maximal differential treatment effect. The SIDES algorithm instead allows the data to guide the process of identifying the subgroups, thus eliminating the need to predefine potential cutoff value. Insights suggested by data in this manner can provide insights that might not be evident on an a priori basis. This would provide a unique advantage for exploratory analyses, for which the primary interest is often to discover new evidence instead of confirming a priori hypotheses. The SIDES algorithm was implemented by a subgroup identification tool (SIT). This tool was developed by Eli Lilly and Company and was used to support the refinement of clinical development and post hoc analyses of study results [9]. Few studies using the SIDES method to identify subgroups of interest have already been published [12‐15].

For the purposes of this analysis, the SIDES algorithm search criteria incorporated stratification by baseline HbA1c to control for the differences between patients in baseline HbA1c because it was well accepted that baseline HbA1c would have a direct effect on the change in HbA1c from baseline. Without this stratification, the search may likely end up with subgroups that have the largest difference in baseline HbA1c. The stratification levels used were quintiles estimated by the SIT for the distribution of baseline HbA1c for the 182 patients included in the analysis dataset. The variable importance score was calculated as the sum of the negative natural log of the splitting criterion [9] over all explored subgroups where the covariate was one of the splitters. It may be viewed as a measure of the relative importance; the covariate contributed to the subgroup identification process. It also provided a benchmark value computed using permutation-based maximum variable importance. To perform the permutation, first, the treatment labels for the 182 at-goal patients were randomly permuted. Second, new analyses using the SIDES algorithm were performed using the same set of covariates, and the maximum variable importance was recorded (this maximum variable importance may correspond to different covariate at different runs). These two steps were repeated 500 times which yielded 500 maximum variable importance scores. We then calculated the mean and standard deviation (SD) of those 500 scores and used them as a benchmark value. Covariates with variable importance greater than the mean of permutation-based maximum variable importance may be considered a strong indication for an important subgroup effect, and covariates with variable importance greater than the mean plus SD of permutation-based maximum variable importance may be considered a very strong indication of an important subgroup effect. The detailed statistical method was described in detail in Lipkovich et al. [9] and Hardin et al. [12].

All p values were provided as a descriptive measure for the difference between treatments and were computed using a two-sample t test for continuous variables and Fisher’s exact test for categorical variables. Such summaries and comparisons were made between PPT and BBT treatment groups in the overall at-goal patients and in the identified subgroup.

Note that SIDES is just a data-mining algorithm that can efficiently explore the features of the observed data and provide a data-driven subgroup of interest. It does not alter the observed data, and the summaries and comparisons are generated using the common and conventional statistical tools.

The outcome variable input into the SIT was change from baseline in HbA1c for patients who achieved control at 24 weeks [<7.0% (53.0 mmol/mol)]. This measure was used as the response variable for the subgroup identification analysis with the SIT, and the following variables were used as candidate predictors of treatment effect: baseline FBG (from self-monitored blood glucose), race/ethnicity, baseline morning 2-h postprandial glucose (from self-monitored blood glucose), duration of diabetes, age at study entry, gender, baseline BMI, baseline use of thiazolidinedione (Yes or No), baseline use of metformin (Yes or No), and randomized treatment. The SIT compared this change in HbA1c between the two groups to identify the subgroup that maximized the treatment difference (PPT versus BBT).