Predictors of treatment response to liraglutide in type 2 diabetes in a real-world setting

Aims There is an unmet need among healthcare providers to identify subgroups of patients with type 2 diabetes who are most likely to respond to treatment. Methods Data were taken from electronic medical records of participants of an observational, retrospective study in Italy. We used logistic regression models to assess the odds of achieving glycated haemoglobin (HbA1c) reduction ≥ 1.0% point after 12-month treatment with liraglutide (primary endpoint), according to various patient-related factors. RECursive Partitioning and AMalgamation (RECPAM) analysis was used to identify distinct homogeneous patient subgroups with different odds of achieving the primary endpoint. Results Data from 1325 patients were included, of which 577 (43.5%) achieved HbA1c reduction ≥ 1.0% point (10.9 mmol/mol) after 12 months. Logistic regression showed that for each additional 1% HbA1c at baseline, the odds of reaching this endpoint were increased 3.5 times (95% CI: 2.90–4.32). By use of RECPAM analysis, five distinct responder subgroups were identified, with baseline HbA1c and diabetes duration as the two splitting variables. Patients in the most poorly controlled subgroup (RECPAM Class 1, mean baseline HbA1c > 9.1% [76 mmol/mol]) had a 28-fold higher odds of reaching the endpoint versus patients in the best-controlled group (mean baseline HbA1c ≤ 7.5% [58 mmol/mol]). Mean HbA1c reduction from baseline was as large as − 2.2% (24 mol/mol) in the former versus − 0.1% (1.1 mmol/mol) in the latter. Mean weight reduction ranged from 2.5 to 4.3 kg across RECPAM subgroups. Conclusions Glycaemic response to liraglutide is largely driven by baseline HbA1c levels and, to a lesser extent, by diabetes duration.


Introduction
Liraglutide is a once-daily human glucagon-like peptide-1 (GLP-1) analogue available for the treatment of type 2 diabetes (T2D), and its efficacy and safety have been demonstrated in the Liraglutide Effect and Action in Diabetes (LEAD) study programme [1][2][3][4][5][6][7]. Liraglutide has also cardioprotective benefits in patients with T2D at increased risk of cardiovascular disease [8]. Liraglutide was approved in the EU in 2009, and data from real-world observational studies have further demonstrated that the benefits of liraglutide on glycated haemoglobin (HbA 1c ) and body weight loss were consistent with those obtained in the randomised LEAD trials [9]. Long-term studies indicated that the benefits were sustained for up to 3 years [10,11].
Liraglutide has been demonstrated to have benefits across a diverse spectrum of patients with T2D, but the extent of HbA 1c improvement differs within patient groups having different demographics and clinical characteristics [12]. Thus, there is an unmet need to identify subgroups of patients with T2D receiving liraglutide who are most likely to have the greatest response to treatment. This information would help healthcare providers individualise treatment options and assess cost benefits. Patients and healthcare professionals could benefit from a more detailed understanding of factors associated with improved response to liraglutide.
The ReaL study (ClinicalTrials.gov identifier: NCT02255266) was the largest observational study of liraglutide in Italian clinical practice, showing that 43.5% of patients achieved HbA 1c reduction ≥ 1% (10.9 mmol/mol) after 12 months of treatment (primary endpoint). This manuscript reports findings from a secondary analysis performed to identify subgroups or classes of patients with T2D who were more likely to have an improved response to liraglutide owing to specific combinations of clinical and socio-demographic characteristics.

Materials and methods
ReaL was an observational, retrospective, longitudinal, multicentre study involving 45 Italian diabetes clinics throughout the country. The design and methods of this real-world study have been previously reported [13]. Briefly, all consecutive patients aged ≥ 18 years diagnosed with T2D and receiving their first prescription of liraglutide in 2011 were eligible for the study. This study was conducted in accordance with the Declaration of Helsinki (last amended by 59th WMA General Assembly, Seoul, October 2013) and the Guidelines for Good Pharmacoepidemiology Practices (ICH-GPP Revision 2, April, 2007). A written informed consent, approved by an independent ethics committee, was signed by all patients before data collection. Data on a range of key clinical variables were obtained from electronic medical records. Information on fasting plasma glucose (FPG), body weight, body mass index (BMI), diabetes duration, presence of diabetes complications, liraglutide treatment, and treatment with other oral antidiabetic drugs (OADs) was extracted at the date of the first liraglutide prescription at baseline in 2011 and after 12 months. The frequency of patients achieving HbA 1c reduction ≥ 1% (10.9 mmol/ mol) after 12 months' treatment (primary endpoint) was calculated. This primary endpoint was selected because it represents a mean effect seen in randomised clinical trials of liraglutide and is a strong indicator of effectiveness that is meaningful to both patients and clinicians. It is also in line with the trend in clinical care to individualise specific HbA 1c targets. Information on side effects and adverse events was not explored, since it was not available in the electronic medical records in a standardised format.

Statistical analysis
Results are expressed as mean and standard deviation (SD) for continuous variables, and proportion and percentages for categorical measures, respectively. Between-group patient characteristics were compared with a Mann-Whitney U test or Student's t test (as appropriate) for continuous variables, or a Chi-square test for categorical variables. Univariate logistic regression was used to identify baseline characteristics of patients who achieved the primary endpoint (HbA 1c reduction ≥ 1.0% [10.9 mmol/mol] at 12 months), compared with those who did not.
Multivariate logistic regression analysis was performed to identify independent factors associated with the endpoint after adjustment for other variables. Covariates included in the multivariate analysis were age, sex, diabetes duration, baseline HbA 1c , FPG, BMI, presence of diabetes complications, treatment at the first prescription of liraglutide (baseline), treatment modality, liraglutide dose, hypertension, dyslipidaemia, and estimated glomerular filtration rate (eGFR) levels. Standardised criteria which were used for diagnosis of hypertension were not established a priori for this study. Data were collected from electronic medical records, but in the Italian national guidelines, hypertension and dyslipidaemia cut-offs are blood pressure (BP) values ≥ 140/90 mmHg and low-density lipoprotein (LDL)-cholesterol ≥ 100 mg/dl, respectively. Covariates used in the multivariate analysis were chosen based on clinical judgment and did not depend on reaching statistical significance in the univariate analysis.
Results are shown as odds ratios (ORs) and 95% confidence intervals (CI).
RECursive Partitioning and AMalgamation (RECPAM) analysis, a tree-based statistical method that integrates standard regression and tree-growing techniques, was used to detect potential interactions among the different variables in predicting reduction of at least 1% in HbA 1c and identify homogeneous and distinct subgroups of patients with increased likelihood of reaching the endpoint [14]. In diabetes, RECPAM analysis has been previously used to identify: patients with T2D at risk of microalbuminuria [15], factors associated with impaired quality of life in patients using continuous subcutaneous insulin infusion [16], and patients at higher risk of cardiovascular disease [17]. The RECPAM analysis was performed using SAS ® (Release 9.4 Cary, NC, USA) and a macro-routine written by F. Pellegrini and updated by M. Scardapane and G. Lucisano. At each partitioning step, the RECPAM method automatically chose the covariate and best binary split to maximise the difference in risk of experiencing the outcome. The algorithm stopped when user-defined stopping rules were met. In this case, each final class was required to have at least 100 patients in total and 30 patients with the target endpoint.
The set of variables tested in the RECPAM analysis was the same tested in the multivariate logistic regression analysis. Continuous variables were not categorised so as to allow the algorithm to choose the natural cut-off points when identifying distinct subgroups of patients. For each subgroup or class, the proportions (%) of patients reaching the endpoint and the likelihood (ORs and 95% CI) to reach the endpoint versus the reference subgroup were obtained. Finally, to detect additional global correlates (i.e. variables playing a role for all patients, irrespective of the interactions detected by RECPAM), a logistic regression model with RECPAMidentified subgroups and all the covariates ruled out by the algorithm was performed. No imputation was used for missing data, and sensitivity analyses were not performed.

Results
A total of 1723 patients were included in the analysis. Baseline characteristics, including diabetes complications and prior treatment regimens, are shown in Table 1. At baseline, most patients were being treated with metformin, either as monotherapy (n = 803, 46.6%) or with sulphonylureas (n = 457, 26.5%). Few patients (n = 100, 5.8%) received insulin. Most patients received liraglutide as an add-on to previous therapies (63.2%), with 33.4% replacing another prior drug with liraglutide, and 3.4% reducing the number of prior therapies. Mean BMI at baseline was 35.6 ± 5.9 kg/m 2 , with 83.3% of patients considered to have obesity (BMI > 30 kg/ m 2 ). By 12 months (primary endpoint analysis), a total of 194/1723 (11.2%) patients had discontinued liraglutide treatment. For those with a known reason (n = 166), most (n = 75/166) were owing to lack of effectiveness. An additional 35 discontinued due to liraglutide intolerance, 28 owing to gastrointestinal side effects, and 20 discontinued for other reasons. A total of 19 patients were non-adherent to therapy. At 12 months, there were 1325 (76.9%) patients with HbA 1c values available at both baseline and 12 months, and 577/1325 (43.5%) reached the primary endpoint (HbA 1c reduction ≥ 1.0% [10.9 mmol/mol]).

Logistic regression analysis
The odds of achieving the primary endpoint, by patient characteristic, are shown in Table 2. In the univariate analysis, higher HbA 1c at baseline was associated with significantly higher odds (OR 2.78; 95% CI [2.43; 3.18]; p < 0.0001). Shorter diabetes duration was associated with a significantly lower odds of reaching the endpoint (OR 0.98; 95% CI [0.97; 1.00]; p = 0.04). Higher diastolic BP (OR 1.02; 95% CI [1.01; 1.04]; p = 0.0002) and higher total cholesterol (OR 1.00; 95% CI [1.00; 1.01]; p = 0.0203) were also associated with significantly increased odds of reaching the endpoint. Other patient characteristics, such as age, sex, BMI, presence of various diabetes complications, dyslipidaemia or eGFR levels, were not significantly associated with odds of reaching the endpoint.
Prior treatment (including insulin) was not significantly associated with reaching the primary endpoint (p > 0.05). However, after adjusting for potential confounding in the multivariate analysis, all prior treatment regimens (except for other dual therapy, p = 0.06) were associated with a significantly lower odds of achieving the endpoint compared with metformin monotherapy ( Table 2). Regarding treatment modality, patients who had liraglutide added to their prior therapy had a significantly higher odds of achieving the primary endpoint (OR 1.74 95% CI [1.38; 1 3 2.20]; p < 0.0001) compared with patients who switched to liraglutide from their previous therapy. Those results were confirmed in the multivariate analysis.
The proportion of patients using liraglutide at higher doses increased with successive follow-up, with over a third (36.1%) using 1.8 mg at 12 months compared to 5.3% at

Discussion
This is the first RECPAM analysis to identify distinct groups of patients with T2D who were prescribed liraglutide in routine clinical practice according to their predicted degree of response to liraglutide treatment. These data can improve clinical practice by providing a deeper knowledge of factors influencing liraglutide's impact on metabolic control. The key message of this analysis is that only baseline HbA 1c and to a lesser extent diabetes duration were predictive of liraglutide effectiveness. Furthermore, these results for the first time clarify that HbA 1c reduction can exceed 2.0% when baseline levels are > 9.0%. This finding has important clinical and health policy implications for the Italian Drugs Agency (AIFA) regulations, considering that patients with HbA 1c ≥ 8.5% are currently excluded from the GLP-1 receptor agonists' reimbursement policy, which requires HbA 1c between 7.5 (58 mmol/mol) and 8.5% (69 mmol/mol) (AIFA regulations). Different patterns have been reported in clinical trials with regard to dose response with liraglutide. In this study, patients using the 1.2-mg liraglutide dose as maintenance dose were more likely to reach the primary endpoint than those using the higher maintenance dose (1.8 mg). This is likely due to an indication bias because patients struggling to achieve good glycaemic control were up-titrated to the higher dose, but owing to their disease severity, they still did not respond as well as healthier patients who did not require an increased dose. Escalation from the starting liraglutide dose of 0.6-1.2 mg likely occurred earlier after initiation, whereas when escalation to 1.8 mg occurred, it tended to be later in the study.
In line with existing findings [18][19][20], we found that the higher the baseline HbA 1c level, the higher the reduction achieved. Multivariate analysis showed that the likelihood of reaching the endpoint increased by 3.5 times for every 1% HbA 1c increase at baseline. In addition, by applying the RECPAM analysis, the study showed that the likelihood of reaching the endpoint was 28 times higher with baseline HbA 1c > 9.1% as compared to baseline levels < 7.5%. In the EVIDENCE study [21], conducted in France by general practitioners and specialists, on 2029 patients, there was a mean (± SD) HbA 1c reduction from baseline of 1.01 ± 1.54% (from 8.46 ± 1.46 to 7.44% ± 1.20; p < 0.0001); after 2 years, 29.9% (95% CI 27.7; 31.2) of patients still had HbA 1c ≤ 7.0%; in the cohort treated within specialist care settings (N = 1398), HbA 1c reduction was − 0.8%.
In the current study, although there were differences in the degree of liraglutide response, each RECPAM class showed decreases in HbA 1c from baseline after 12 months of treatment. As might be expected, a greater proportion of patients with the poorest glycaemic control at baseline achieved the primary endpoint of HbA 1c reduction ≥ 1.0% (10.9 mmol/ mol) after 12 months, since it would be incrementally more difficult to achieve that degree of absolute HbA 1c reduction in patients already at or near glycaemic targets. Nevertheless, these results suggest that there is a distinct subgroup of patients for whom liraglutide treatment can help achieve HbA 1c reductions in excess of 2.0% (21.9 mmol/mol), a finding that may have important clinical implications.
The RECPAM algorithm selected only baseline HbA 1c and diabetes duration as important splitting variables when creating the responder subgroups or classes. This indicated that other patient variables were less important in determining the degree of response to liraglutide. Although BMI was not selected by the algorithm, this too may be because of the high prevalence of obesity in the sample. Multivariate logistic regression with RECPAM categories forced into the model further confirmed that liraglutide is best used as an add-on to, rather than replacement for, prior treatment regimens (generally OADs) in T2D (OR 1.79; 95% CI [1.29; 2.50]). This finding is in line with current treatment guidelines [22]. Interestingly, the largest patient subgroup (n = 306, RECPAM Class 5) (Table 3) had comparatively good HbA 1c control (≤ 7.5% [58 mmol/ mol]), suggesting that there is also a patient subgroup who may initiate liraglutide to pair the glycaemic control to weight loss.
Regarding the role of diabetes duration, a previous study on liraglutide reported a higher efficacy in patients with short diabetes duration [12], while the ReaL study [13] found improvements in metabolic control also in patients with long diabetes duration. RECPAM analysis clarifies that diabetes duration can play a role mainly for patients with HbA 1c levels between 7.5 and 8.2%; in particular, one in two patients with diabetes duration ≤ 5 years reached the endpoint, compared to one in three for a diabetes duration > 5 years. The role of BMI and previous therapy as independent predictors emerging in other studies [19,23] was not confirmed in our study.
A strength of this study was the large sample size. Use of real-world data also makes the findings more generalisable to patient populations seen in regular clinical practice. The observational nature of the study may introduce bias in the selection of patients who were prescribed liraglutide; however, consecutive enrolment of all patients was adopted to minimise this. Since these results reflect the clinical usage of liraglutide in Italy, they may not be generalisable to countries with different usage patterns. As a retrospective study based on electronic medical records, the completeness of information depended on the ability of participating centres to record clinical data. It should be noted that data completeness was judged satisfactory (i.e. 97.2-56.3% complete for the adjustment variables used). Insulin secretion capacity was not evaluated as a potential predictor of HbA 1c reduction with liraglutide, although several studies have suggested the usefulness of this parameter in predicting the effectiveness of liraglutide [24,25]. This would be useful to explore in future studies. We cannot exclude the involvement of other factors, besides HbA 1c and partly diabetes duration, in determining HbA 1c reduction through liraglutide, but we analysed all factors easily available to diabetologists to guide routine clinical practice.
In conclusion, in this study, glycaemic response to liraglutide was largely driven by baseline HbA 1c levels and to a lesser extent by diabetes duration. The clinical benefit seems to be maximised when used as an add-on to prior therapies. All RECPAM classes showed weight loss, which appeared independent of mean HbA 1c reduction. RECPAM analyses suggest an urgent need to revise the AIFA criteria for reimbursement due to the finding that HbA 1c reduction can exceed 2.0% in people with HbA 1c > 9.0%.