Study design and population
This longitudinal retrospective study analyzed data on patients with type 2 diabetes routinely followed at the diabetes outpatient clinic of the University Hospital of Padua from 2008 to 2018. Detailed data on demographics, medical history, and medications were linked with administrative databases including hospital discharge codes and death certificates, as described before [14].
We retained only patients with at least 3 visits and at least 6 months of observation in the database. A schematic representation of entry visit, index date, follow-up time, and end-of-follow up is graphically represented in Additional file 1: Figure S1. The first available visit in the dataset between January 1st 2008 and June 1st 2018 was considered as the entry visit (visit 1). All subsequent visits were included in the evaluation of the exposure. To allow a minimum time between exposure and outcome, outcomes were ascertained after visit 3, which was considered the index date. The end of follow-up was September 1st 2018 (the last available date in the administrative database including death certificates and hospital discharge codes) or the last day the patient was present in the database (e.g. if a patient moved to another region) or the date of death, whichever occurred first. As patients accessed the clinic at least once a year, they were considered lost to follow-up and censored after two years without an outpatient visit or hospital admission.
Individualized patient’s target HbA1c level was calculated according to a previously defined algorithm, using five objective parameters: life expectancy, comorbidities, macro-vascular and advanced micro-vascular complications, risk of hypoglycemia from treatment, and disease duration [15].
Definition of exposure
To evaluate whether treatment was aligned with a modern treatment approach, we identified six domains reflecting the main ADA/EASD recommendations on the management of hyperglycemia in patients with type 2 diabetes issued in 2018 [2], which was the first consensus document embracing a treat-to-benefit rather than a treat-to-target approach. These domains were selected to cover appropriateness of treatments in various stages of disease (Table 1). At each visit, we evaluated, as a binary variable, the following domains to establish whether treatment was aligned with recommendations. Domain 1 (metformin use) was met if first-line treatment included metformin, while it was not when any other combination of treatment not including metformin was used as first-line without having tried metformin before, except for patients with CKD. Domain 2 (intensification) was not met when there was a lack of treatment intensification (i.e., add-on or switch to a different regimen) with an HbA1c was above individualized target for two consecutive visits. Domain 3 (second-line treatment): was not met when sulfonylureas (SU) or insulin were used as second-line treatments. Domain 4 (insulin use) was not met in the presence of an inappropriate use of insulin, defined as initiation of insulin before GLP1-RAs and before metformin, or using bolus regimens before basal insulin. Domain 5 (use of cardioprotective drugs) was not met when cardio-protective treatments (SGLT2is and GLP1-RAs) were not prescribed to patients with previous cardiovascular events or revascularization (without having tried them before). Domain 6 (use of weight-affecting drugs) was not met when weight-increasing drugs (SU, insulin, glitazones) were used in patients with obesity (BMI > 30 kg/m2) prior to weight-neutral or weight-decreasing drugs (metformin, GLP1-RA, SGLT2i, DPP4i). All these 6 domains were evaluated at each visit in each patient.
Patients were in Group 1 when at least one domain was not met (i.e. deviated from recommendations) or in Group 2 otherwise (i.e. when all domains were always met according to recommendations). We compared outcomes of patients in Group 2 versus Group 1. The proportion of time being in Group 2 (cumulative time aligned with recommendations) was evaluated as the ratio between the cumulative months being in Group 2 and the total follow-up time to that visit.
These domains were built according to 2018 ADA/EASD guidelines but can be considered valid also according to the most recent 2022 ADA/EASD guidelines [2, 3]. The main exception is that SGLT2i or GLP1RA can now be considered appropriate as first-line treatment before metformin for patients at high or very high risk of cardiovascular disease, including heart failure, or renal disease. However, when data for this study were collected (2008–2018), such practice was very uncommon and applying a similar domain would made all subjects become unaligned to the recommendation because most patients in tertiary-referral outpatient diabetes centers in Italy have very-high cardiovascular risk [16].
Definition of outcomes
The primary outcome was occurrence of the 3-point major adverse cardiovascular events (MACE, defined as cardiovascular death, non-fatal myocardial infarction, or non-fatal stroke). Secondary outcomes were (i) a composite of hospitalization of heart failure and cardiovascular death (HF-CVM); (ii) all-cause mortality.
Occurrence of the outcomes was ascertained through hospital discharge codes (based on ICD-9) and death certificates (based on ICD-10) as reported in the administrative databases. MACE was defined in the presence of any of the following ICD-9 codes: 410.x (acute myocardial infarction) or 430, 431, 432.x, 433.x, 434.x, 436, (hemorrhagic or ischemic strokes), or death with the following ICD10 codes (I20-I25, I46). HF-CVM was defined in the presence of any of the following ICD-9 codes: 428.x or death with the following ICD10 codes (I20-I25, I46).
Statistical analysis
Continuous data are presented as mean and standard deviation, whereas categorical variables are shown as percentage. Comparisons between patients in the two groups were performed using Student’s t test or chi squared tests, as appropriate. Cox regression models for time-dependent covariates were used to evaluate the association between Group 2 and MACE, HF-CVM (both including recurrent events), or all-cause mortality. We also used multivariable adjusted models with increasing complexity to account for possible “healthy users bias” due to patients in generally better health status (e.g. on first-line treatment, or without obesity or with HbA1c at target) being more likely to be in Group 2. Model 1 included age, sex, study entry year, diabetes duration, and the following time-varying covariates: presence of cardiovascular disease (myocardial infarction, ischemic myocardial disease, cardiac revascularization or stroke), diabetic kidney disease (defined by reduced eGFR below 60 ml/min/1.73 m2 and/or albuminuria), macrovascular disease (including history of cardiovascular events and clinical or subclinical peripheral artery disease, e.g., presence of ultrasound-detected carotid plaque), and microvascular disease (retinopathy, neuropathy, or nephropathy). Model 2 was similar to model 1 with the addition of the following time-varying variables: line of treatment, latest BMI and HbA1c values, other medications (antiplatelet, statins, other lipid-lowering treatments, RAS blockers, calcium channel blockers, beta-blockers, diuretics, oral anticoagulants), presence of obesity and severe diabetes decompensation (HbA1c levels higher than 10%; yes/no). Model 3 was similar to model 2 with the addition of the history of cancer, chronic obstructive pulmonary disease, systemic inflammatory disease, and ultrasound-documented hepatic steatosis. Sensitivity analyses stratifying on cumulative time being in Group 2 greater or less than 50% were conducted using the fully adjusted model (model 3). The analyses on HF-CVM included also history of baseline HF in all models. The impact of lifestyle (i.e. smoking habits, alcohol consumptions and physical activities) or and socio-demographic variables (i.e. level of instruction, citizenship, marriage status) on top of model 3 was evaluated in additional sensitivity analyses, since these information were available for a subset of patients.
We also stratified patients according to key baseline characteristics to evaluate whether the impact of exposure on the outcomes was affected by the patients’ clinical phenotype. We performed additional sensitivity analyses using Cox Marginal Structural Models (MSM, implemented via the SAS %MSM macro), by fitting a weighted pooled logistic model using inverse probability weights for treatment and censoring. Briefly, by an inverse-probability of treatment weighing evaluated at follow-up visits, these Cox MSMs allow for appropriate adjustment of confounding when there are time-dependent confounders that might themselves be affected by previous treatment or exposures, e.g. accounting for confounding by indication and healthy users biases [17,18,19,20].
The impact of each domain on the overall effect of Group 2 on MACE, HF-CVM, and mortality was tested by removing one domain at a time. This defined 6 alternative Group 2 definitions, each excluding one of the six domains. Then, the association of these alternative Group 2 definitions with outcomes was tested, and their estimates compared by Wald test to those obtained using the standard definition of Group 2.
Extraction of electronic medical records allowed complete collection of data on medication prescriptions (with no missing). According to study diagram described in Additional file 1: Figure S2, all subjects included in the analyses had information allowing complete evaluation of all domains and all covariates used in the different multivariable adjusted models (all models were tested on the same number of subjects). The only exception was the sensitivity analyses adjusted by lifestyle and socio-demographic information that were available only in a subset of individuals. Statistical analysis was performed with SAS and significance set to p < 0.05.