Data source and study population selection

This is a retrospective longitudinal cohort study. This study was carried out based on data sourced from Nivel Primary Care Database (Nivel-PCD). Nivel-PCD systematically collects data from approximately 500 general practices in the Netherlands (around 10% of the Dutch population) including prescribed medications recorded by general practitioners (GPs) [18]. Hence, Nivel-PCD contains GPs’ prescribing records, namely medications ordered by the GPs within the Dutch health system. Pharmacy dispensing/claims data is not captured in Nivel-PCD. In Dutch primary care, long-term therapies are issued as repeat prescriptions with an authorisation typically valid for up to 12 months. Community pharmacies generally dispense 1–3 per month quantities per issue. Insulins and oral antidiabetic drugs (OADs) follow the same framework. Each prescription record in Nivel-PCD includes the calendar date, Anatomical Therapeutic Chemical Classification (ATC) code, prescribed daily dose/instructions, and intended duration or quantity. These fields were used to derive days of coverage per prescription and to build 12-month prescription histories from the index date (first OAD prescription). For this study, these data were integrated with patient characteristics, healthcare consultations, patient morbidity profiles, laboratory results, and the primary or secondary healthcare provider managing diabetes care. In the Netherlands, type 2 diabetes accounts for 82% of all diabetes cases [19], with negative consequences for public health [20–22]. The age and gender distribution of patients registered in Nivel-PCD reflects those of the general Dutch population with type 2 diabetes [23]. Inclusion criteria included all patients first diagnosed with T2D (identified by the International Classification of Primary Care [ICPC] code T90) between 2015 and 2019 within the Nivel PCD data source. Index date was set as the date of the first prescription of an oral antidiabetic (OAD) medication. OADs were identified in the Nivel-PCD by the ATC code A10B, as follows: biguanides (A10BA), sulfonylureas (A10BB and A10BC), combinations of oral blood glucose lowering drugs (A10BD), alpha glucosidase inhibitors (A10BF), thiazolidinediones (A10BG), DPP-4 inhibitors (A10BH), SGLT2 inhibitors (A10BK) and other blood glucose lowering drugs, excluding insulins (A10BX). The OAD prescribed at the index date was defined as the index drug. Medication adherence for each patient in the cohort was calculated with reference to this index drug. Each patient was followed for 12 months from index (one-year follow-up). Patients with any insulin prescription before the index date were excluded. Baseline cohort characteristics were observed within the 90 days prior to the index date for each patient. All baseline covariates were assessed within this window (e.g., age, sex, comorbidity measures, concomitant chronic medicines/diagnoses). For each patient, follow-up started at the index date and continued for 12 months (sliding one year follow-up window per patient). Observation period was censored on 31 December 2020. Patients who discontinued the index OAD prescription before 12 months were not excluded. Their discontinuation contributed to the persistence endpoint, while they continued to contribute on outcome data within the observation window (S1 Fig). Insulin addition was evaluated for each patient and defined as the first insulin prescription initiated after the index date within the follow-up.

Trajectory modelling and medication adherence estimation

Medication adherence assessment was performed for each patient within the 12-month prescription history (one-year follow-up) starting from the index date. This single time series (days 0–365) was the basis for all adherence measures. Medication adherence was computed according to the EMERGE guidelines [5], distinguishing the three different phases of the adherence process. Initiation phase was calculated as the occurrence of at least two prescriptions of the index oral antidiabetic drug (OAD) within the one-year follow-up period, since dispensing data were not available. Thus, if a second OAD prescription occurred after the first (index) OAD prescription, the patient was classified as an initiator, and the index date was considered as the initiation date. Persistence phase was computerised from prescribing patterns of OADs, considering the prescribed coverage period of each individual OAD. Patients were classed as non-persistent if no later OAD prescription was recorded, or if the gap to the next prescription was > 2 × the expected days’ supply of the previous one. This approach identifies gaps in treatment that indicate discontinuation during the follow-up period. If a patient switched to another OAD and no further doses of the reference drug were prescribed, the reference therapy was considered discontinued at the first permitted break in the therapeutic interval. When another OAD was added and prescriptions for the index drug continued, the index trajectory was unchanged. Patients with only one OAD prescription during follow-up were reported under initiation but excluded from persistence and implementation estimations as, according to the EMERGE guidelines, persistence is time-to-event. This is because these two phases require ≥2 prescriptions to observe continued exposure and to avoid misclassification in the trajectories. Indeed, as persistence calculation is time-to-event, it is not possible to determine whether the only one OAD prescription was taken by the patient or not (e.g., the patient does not start the prescribed treatment). Implementation phase was calculated and plotted by using the trajectory-based modelling, mainly based on clustering approach. Clustering method used consisted of a set of multivariate techniques aimed at recognising and grouping homogeneous items in a data set by using non-parametric statistics [10,24–26]. The clustering approach grouped patients with similar month-to-month adherence rates over the first year of the OAD-therapy. Each cluster was summarized by a centroid, i.e., the typical adherence curve formed by averaging the profiles of its members. Then, the clusters were labelled by the shape of their centroid to reflect recognizable behaviours in adherence. Hence, cluster analysis for adherence implementation measurement allowed the identification of groups of T2D patients with common characteristics (e.g., potential determinants or predictors of risk) [14,24–27]. This approach linked heterogeneous individual adherence patterns into a small set of clinically interpretable behaviours. The indicator used for the adherence calculation in its implementation was the continuous multiple interval measure of medication availability/gaps (CMA) version 9 [28–30]. For the whole observation period, the CMA9 was computed as the ratio of theoretical medication use days to the adherence assessment period duration, accounting for prescriptions carryover and excluding end-period prescriptions. CMA9 provides an average adherence measure over time. It weights days by each prescription’s coverage and calculates adherence in non-overlapping one-month windows, aligned with the typical interval between successive prescriptions. Hence, prescriptions histories for each index OAD over an observation period one year were calculated. In Dutch primary care, chronic therapies are authorised as repeat prescriptions by the GP, commonly for up to 12 months. Community pharmacies then dispense smaller quantities according to standard practice (often 1–3 months at a time for both insulins and OADs [27,31]). Nivel-PCD records the GP prescription authorisations and dates. Hence, medication coverage was computed using successive prescription records. Specifically, days of supply were derived at the prescription level from the recorded prescription duration (days) and the prescribed daily dose. When this field was unavailable or inconsistent, days of supply were inferred from molecule/formulation-specific dosing (national guidance [27]) and the spacing between successive prescriptions. Coverage was carried over across prescriptions and summarized in one-month sliding windows for CMA9. Then, the actual doses prescribed (dosage and coverage) were compared with the doses recommended by the NHG guideline for diabetes mellitus type 2 [27] and the Medicines Information Bank managed by the Dutch Medicines Evaluation Board (MEB) [31]. NHG recommended posology related to OADs are: once to twice daily for sulfonylureas, thiazolidinediones, DPP-4 inhibitors and SGLT2 inhibitors; 1–3 times a day for combinations of oral blood glucose lowering drugs, alpha glucosidase inhibitors, other blood glucose lowering drugs (excluding insulins); 2–3 times a day for biguanides considering that the extended-release tablet is taken once daily. Therefore, the posology of all OAD medications was settled ad-hoc per each medicinal speciality.

Model covariates

Baseline characteristics and post-index events, such as insulin initiation, were analysed for each adherence group. Patients’ complexity was assessed calculating: i) rates of multimorbidity (two or more chronic conditions at baseline based on ICPC codes) and polypharmacy (concurrent use of five or more different medications at baseline based on ATC codes); ii) index of high risk T2D patients according to the Dutch treatment and diagnostic guidelines for T2D [27]: patient who also suffer or suffered from angina pectoris, acute myocardial infarction, other and chronic ischemic heart disease, heart failure, stroke/cerebrovascular accident, atherosclerosis, other arterial obstruction, vascular disease, chronic alcohol abuse, obesity (BMI > 30) and symptoms/complaints kidney; iii) Charlson comorbidity (CCI) score [26]: the score assigns weights for a number of major conditions included in secondary diagnoses. The CCI score appears as the total value of the assigned weights and pertaining measure of the weight of each comorbidity. CCI scores were categorized into three distinct grades: low, characterized by CCI scores of 0–1; mild, corresponding to CCI scores of 2–3; and severe, indicated by CCI scores of ≥4.

Clinical outcomes

Association between medication adherence and main T2D clinical outcomes was investigated in accordance with the guideline of the Dutch College of General Practitioners (NHG) for T2D monitoring [27]. The guideline suggests that T2D patients must undergo quarterly check-ups aimed at evaluating potential diabetes-related complications. These check-ups include the assessment of specific clinical parameters against clinical target levels according to WHO guidelines. The main clinical parameter includes haemoglobin A1c levels (HbA1c; clinical target ≤53 mmol/mol/ ≤ 7%). Also, secondary outcomes are LDL cholesterol levels (target <2.8 mmol/L), body weight measured through Body Mass Index (BMI; target <27), and blood pressure (BP; target <140/90 mmHg). These latter were treated as secondary, descriptive outcomes. Hence, clinical parameters were observed for each T2D patient within 90 days before the index date and within 90 days after the one-year follow-up period (S1 Fig). Thus, the end of the entire study period was on 31 March 2021. If the T2D patient had several clinical parameter measurements before the index prescription or after the end of follow-up, the measurement closest in time to these two set dates was considered. Additionally, number of clinical measurements collected for each T2D patient within the one-year follow-up period was also retrieved to assess temporal variation and compare trends across adherence groups.

Statistical analyses

Covariates were described using the mean and standard deviation (SD) or median and interquartile range (IQR) for continuous variables, and frequency and percentage for categorical variables. Adherence trajectory groups were treated as the exposure accumulated during the one-year follow-up. Pre/post changes in clinical parameters by adherence trajectory were compared. Because adherence trajectories and outcomes are derived over the same 12-month horizon, all contrasts were interpreted as associations rather than causal effects. Clustering in adherence groups was performed with the AdhereR package (version 0.8.10) and the R package “kml” (version 2.4.1), which provided an implementation of k-means designed to work specifically on longitudinal data [28–30]. The number of adherence groups was determined by comparing model fit statistics across solutions ranging from two to six groups of adherence (k = 2–6) using one-year CMA9 series. Implementation exposure was constructed as monthly CMA9 values over 12 months (rolling 1-month windows) [24–26]. Solutions with a multi-criteria panel of internal indices (standardized Calinski–Harabasz CH- CH2- CH3, Ray-Turi, Davies-Bouldin) were compared. Cluster stability was assessed in three steps. First, k-means was run with 100 random initialisations to reduce dependence on starting values. Second, bootstrap resampling was applied (200 samples with replacement). For each bootstrap sample, the solution was compared with the full-sample solution using the Jaccard index for cluster membership. Specifically, values ≥0.75 were considered stable and values 0.60–0.75 moderately stable. Third, a split-sample check was performed: centroids were estimated in a random 50% training subset, and the remaining 50% were assigned to the nearest centroid. Finally, agreement with the full-sample assignment was then evaluated. Moreover, sensitivity analyses were performed as follow: 1) the monthly CMA9 construction (±3-day tolerance around month boundaries); 2) application of a light 3-point moving average (yes/no); 3) a minimum cluster size of 5% to avoid small clusters. Finally, the selection rule was the smallest k with jointly favourable indices, stable membership (Jaccard ≥ 0.75), and distinct (also clinically interpretable) trajectory shapes [24–26].

Multivariate logistic regression was performed to identify factors associated with adherence clusters. A stepwise selection procedure was used to select covariates to be included in the models. These covariates included: sex, age category (<25, 26–39, 40–59, 60–79, ≥ 80 years), high-risk T2D status per NHG guidance [27], hypertension at baseline, baseline HbA1c within target (≤53 mmol/mol; 7%), baseline blood pressure within target (<140/90 mmHg), and insulin addition during follow-up (post-index indicator). Results are presented as adjusted odds ratios (aORs) with 95% confidence intervals and two-sided p-values (α = 0.05). Linear mixed-effect regression model was used to model HbA1c over the 12-month follow-up and its association with adherence groups. All HbA1c measurements recorded during follow-up were included. The fixed-effects structure comprised time since index (measured in months), adherence trajectory group, and their interaction (time x group), with baseline HbA1c entered as an adjustment covariate. Random effects included patient-level intercepts and slopes for time, to consider the association within-patient and heterogeneous trajectories. Model coefficients, interaction contrasts, and monthly slopes of HbA1c were reported with 95% confidence intervals. For both models p value <0.05 was considered statistically significant.

Data management was performed with Stata, Statistical software for data science. Statistical analyses were performed with R (version 3.6, The R Formulation for Statistical Computing).

Adherence groups and cohort characteristics

The overall cohort consisted of 3,404 patients first diagnosed with T2D. A flow chart reporting final cohort selection is showed in S2 Fig. Baseline cohort characteristics assessed at the index date are detailed in Table 1 and S1 Table. Regarding the clustering, the solution of k = 4 clusters achieved the near-highest standardized CH indices and favourable Davies-Bouldin. The bootstrap stability for the k = 4 clusters solution was high (cluster-wise Jaccard 0.79–0.88; overall mean 0.84) and split-sample agreement was 89%. Sensitivity analyses produced highly correlated centroids (r = 0.93–0.98) and similar group proportions (absolute differences ≤ 1.8%) (see S3 and S4 Figs). The identified four adherence groups were as follows:

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Patients’ baseline characteristics by each adherence group.

https://doi.org/10.1371/journal.pone.0342056.t001

1. i). Group A: Perfect Adherence n=2,386 T2D patients (70.1%).
2. ii). Group B: Slow decline in adherence n = 453 T2D patients (13.3%).
3. iii). Group C: Low Adherence n=362 T2D patients (10.6%).
4. iv). Group D: Slow increase in adherence n = 203 T2D patients (6.0%).

Overall, the initiation phase analysis revealed that a small proportion of T2D patients of the cohort (0.9%) had a single OAD prescription. However, Group C (low adherence) had the highest proportion of patients with only one OAD prescription (6.9%) (S2 Table). In the persistence phase, using the pre-specified permissible-gap definition, 84.6% of the overall T2D cohort remained on therapy at the end of the one-year follow-up. Specifically, patients in group A (perfect adherence) had the highest persistence rate (98.8%) at the end of the follow-up period, followed by group D (slow increase in adherence, 72.4%), group B (slow decline in adherence, 62%) and group C (low adherence, 26%) (S2 Table).

Regarding the implementation phase, the mean CMA value across the total cohort was 0.8 (SD 0.3) with differences between groups: group A maintained optimal implementation (CMA 1.0), group C recorded very poor implementation (CMA mean 0.3, SD 0.1) and group B and D had intermediate profiles (CMA mean 0.6 and 0.7 respectively, with SD 0.1) (S2 Table). Age differences were identified across the four adherence groups (Table 1). Mainly, about half of patients in the perfect adherence group and slow decline in adherence group were aged between 40–59 years (group A: n = 1,085, 45.5%; group B: n = 202, 44.6%) while half of the patients in the low adherence group and slow increase in adherence group were aged between 60–79 years (group C: n = 163, 45%; group D: n = 93, 45.8%). Moreover, 17.4% of patients with low adherence to OADs treatment and 12.1% of patients with slow decline in adherence added an insulin treatment 2½ months after the start of OADs treatment (S3 Table). The highest proportion of high risk T2D patients was identified in the perfect adherence group (group A, 12.5%). Of these, 26.2% suffered from an acute myocardial infarction, and in the slow increase in adherence group (group D, 11.3%), of these 26.1% had a stroke/cerebrovascular event. Patients in the low adherence group C recorded the highest comorbidity score (mean: 2.6; SD: 11.2). Regarding the overall T2D cohort multimorbidity level, the mean number of comorbid conditions was higher for patients in the slow increase in adherence group (group D) with about three additional chronic conditions to T2D.

Adherence groups and clinical outcomes

Variation in clinical outcomes from the index date to the end of the one-year follow-up recorded differences between the four adherence groups (Table 2, S5 and S6 Figs). The clinical parameters target values to be monitored in patients with T2D are described in S4 Table.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Variation in outcome measures from the beginning to the end of follow-up.

https://doi.org/10.1371/journal.pone.0342056.t002

Prior to the start of OADs treatment (index date), 70.5% of T2D patients with perfect adherence (group A) had measured HbA1c levels >53 mmol/mol, but after one year from the start of therapy (end of follow-up) only 31.5% of these were still above the clinical target. Highest variation was also observed for the group D (slow increase in adherence) recording 61.6% of patients above the HbA1c clinical target at the beginning of follow-up and 25.1% at the end (Table 2, S5 and S6 Figs). Moreover, HbA1c levels at the end of follow-up were measured below the target value for all adherence groups except for the patients in the low adherence (group C 53.7 mmol/mol at the end of follow-up) (see Fig 1 and S5 Table) although they started from above target levels at baseline (S7 Fig). Additionally, the mixed-effects model showed that all groups had a negative monthly slope in HbA1c after OAD initiation (overall −2.93 mmol/mol per month). Compared with the perfect adherence group A, the slope was significantly less steep in group B (difference +0.44 mmol/mol per month; 95% CI 0.08–0.80; p = 0.02) and group C (+1.07; 95% CI 0.61–1.52; p < 0.001), indicating slower HbA1c improvement in these groups (Fig 1 and S6 and S7 Tables).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Linear mixed-effect regression model correlating changes in HbA1c levels and adherence groups.

The regression line reflects the mixed-effects model that uses all HbA1c timestamps and includes time since index date and baseline HbA1c. Each grey dot represents an individual HbA1c measurement recorded during the follow-up for a T2D patient. The black line illustrates the fitted linear mixed-effects trend. The model includes time since index (months), adherence group, and their interaction, adjusts for baseline HbA1c. The model also uses patient-level random intercepts and slopes. The dashed horizontal line represents the clinical HbA1c target threshold, set at 53 mmol/mol (7%).

https://doi.org/10.1371/journal.pone.0342056.g001

Normal LDL levels (<2.8 mmol/L) were only reached by T2D patients in the perfect adherence group (group A). Albeit this, at the end of the follow-up, 38.1% of perfect adherent were still above the LDL clinical target. Regarding BMI parameter, most T2D patients across all adherence groups were overweight at baseline (BMI > 27), especially in the perfect adherence (group A: 62.4%) and slow increase (group D: 59.1%) groups. At the end of follow-up, mean BMI values recorded slightly lower levels in all adherence groups (e.g., from 30.7 to 30.2 in group A).

Blood pressure control recorded lower levels for all the adherence groups from the baseline to the end of follow-up. The proportion of T2D patients with BP > 140/90 mmHg declined in all groups, from 39.6% to 25.4% in the perfect adherence group A. Mean BP values reached similar levels across groups at the end of follow-up (approximately 133/80 mmHg).

Logistic regression models were performed to evaluate the association between determinant variables with the adherence group, as shown in Table 3. Perfect adherence behaviour to OADs treatment was associated with age 40−59 years (OR 2.83, CI 95% 1.12–7.18) diagnosis of hypertension (OR 1.21, CI 95% 1.00–1.47) and with a high-risk T2D (OR 1.35, CI 95% 1.05–1.73). Additionally, perfect adherence was associated to subjects reaching normal values of HbA1c after one-year OADs treatment (OR 1.35, CI 95% 1.14–1.60) as for those adding insulin to the initial OAD treatment after the index date (OR 0.55, CI 95% 0.42–0.71).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Multivariate logistic regression analysis to identify determinants of adherence groups.

https://doi.org/10.1371/journal.pone.0342056.t003

The model also revealed that patients with hypertension and those with normal HbA1c values at the end of follow-up were less likely to belong to the low adherence group (OR 0.7, CI 95% 0.52–0.94 and OR 0.75, CI 95% 0.58–0.96, respectively) and to the slow decline in adherence group (OR 0.76, CI 95% 0.58–0.99 and OR 0.81, CI 95% 0.65–1.02, respectively).