ABSTRACT
Patients' longitudinal adherence to antidiabetic medication in routine clinical care remains unexplored. This study aimed to identify adherence groups among individuals with type 2 diabetes with up to 1 and 5 years of follow‐up. This was a register‐based cohort study using data from Swedish national health and population registers and the National Diabetes Register (2006–2022). New users of blood glucose–lowering drugs (other than insulin) were identified. Trajectories of the proportion of days covered (PDC) by any antidiabetic medication, including insulin, over 1‐ and 5‐year periods were clustered using k‐means for longitudinal data. Analyses up to 1‐ and 5‐year follow‐up periods included 75,421 individuals with an overall mean PDC of 0.7 and 283,795 individuals with an overall mean PDC of 0.3, respectively. K‐means clustering identified two main adherence groups. For the 1‐year follow‐up, 70.6% of individuals were in the cluster with a higher mean PDC (0.9) and 29.4% in the cluster with a lower mean PDC (0.4). The corresponding figures for the 5‐year follow‐up were 36.9% (higher mean PDC [0.9]) and 63.1% (lower mean PDC [0.3]). Clusters with higher mean trajectories of PDC included more men, older individuals, patients using drugs from only one antidiabetic medication class, and noninsulin users during follow‐up. Mean trajectories of adherence decreased mainly during the first year. This study identified a substantial problem with longitudinal adherence to any antidiabetic medication, with a low proportion of individuals clustered as having higher adherence during the 5‐year follow‐up. Results suggest the need for interventions via follow‐up strategies aiming at monitoring and improving continuous treatment management while considering tailored treatment strategies.
Keywords: antidiabetic medication, clusters of adherence, longitudinal trajectories, pharmacoepidemiology, type 2 diabetes
Summary.
- What is the current knowledge on the topic?
-
○Adherence to pharmacological treatment for type 2 diabetes is crucial for disease control, but patients' longitudinal adherence in routine clinical care remains unexplored.
-
○
- What question did this study address?
-
○Which longitudinal trajectories of adherence to antidiabetic medication exist and which patient groups do they include?
-
○
- What does this study add to our knowledge?
-
○This large cohort study investigating longitudinal adherence to antidiabetic medication in routine clinical care identified two main clusters of adherence over a 5‐year follow‐up. Mean trajectory of adherence decreased mainly during the first year, and only 36.9% of individuals were clustered in the high adherence group. This cluster included more men, older individuals, users of only one antidiabetic class, and noninsulin users. Despite some differences in individuals' characteristics, the study showed a substantial problem with adherence for all patient groups, which urges interventions.
-
○
- How might this change clinical pharmacology or translational science?
-
○This study showed that the proportion of individuals with high adherence is low over a 5‐year follow‐up. Targeted interventions could sustain adherence over time via follow‐up visits aiming at the improvement of continuous treatment management. Some differences in characteristics may help to target closer monitoring of patients at risk of low adherence during the initial treatment phase. This may also reduce the risk of low compliance with medication and accelerate the identification of the treatment strategies that could improve long‐term adherence.
-
○
1. Introduction
In 2021, the International Diabetes Federation reported that 537 million adults have diabetes globally, the majority of whom have type 2 diabetes. When lifestyle modifications are not sufficient to reach normal hemoglobin A1c (HbA1c) levels, pharmacological treatment is required and usually initiated with metformin, the most commonly recommended first‐line treatment. Strategies for the treatment of type 2 diabetes are highly individualized and can reach different levels of complexity; that is, metformin can be combined with or substituted by other blood glucose–lowering drugs [1]. Treatment strategies can also change over time due to tolerance to treatment, preferences, cost, and new antidiabetic medication becoming available on the market. Nevertheless, most individuals with type 2 diabetes will need lifelong treatment to prevent or alleviate long‐term complications.
Adherence to lifelong recommended treatment strategies is crucial for proper control of the disease, but is influenced by multiple aspects including individuals' acceptance of the disease and the need for medication [2], as well as strategy complexity, efficacy, tolerability, and route of administration. Moreover, patients' adherence to treatment in routine clinical care is not a static behavior but can vary over time; application of novel methods leveraging national health registers to identify treatment episodes of continuous use of antidiabetic medication [3] can contribute to the identification of longitudinal trajectories of adherence, and knowledge gained could be used, for instance, to inform intervention strategies.
Despite its key role, only a few studies have assessed adherence, and those studies have had a short‐term focus or focused on only one antidiabetic medication class [4]. For example, a cohort study investigating 1‐year adherence to sodium–glucose cotransporter 2 inhibitors (SGLT2i) defined adherence using the proportion of days covered (PDC) based on pharmacy claims [5]. Other studies used drug dispensation data to identify patterns of metformin monotherapy use defined with the PDC over a 1‐year period [6], and the medication possession rate (MPR) to define 90 days of adherence to insulin in patients with type 2 diabetes [7].
Due to a lack of data on long‐term adherence to any antidiabetic medication for type 2 diabetes in routine clinical care, there is an urgent need for observational studies aimed at identifying longitudinal adherence patterns in populations of patients with type 2 diabetes. Built on this literature gap, the present study used longitudinal population‐level data from Swedish national registers and the clustering method K‐means algorithm for longitudinal data [8] to identify treatment episodes of antidiabetic medication use and to cluster individuals with type 2 diabetes with up to 5‐year follow‐up in adherence groups and to describe their characteristics.
2. Material and Methods
2.1. Study Design and Data Sources
This was a register‐based cohort study using data from the Swedish national health and population registers and the National Diabetes Register (NDR). Data were linked via the unique personal identification number given to each Swedish resident at birth to identify the study population and to retrieve information on the characteristics of interest. Specifically, antidiabetic medication use was retrieved from the Swedish Prescribed Drug Register (PDR), which provides information on all prescription drugs dispensed at pharmacies in Sweden since July 2005 [9]. Dates of migration and death were extracted from the Total Population Register and the Swedish Cause of Death Register. Finally, additional information on individuals with diabetes was retrieved from the NDR, launched in 1996 for the purpose of promoting evidence‐based development of diabetes care. The coverage rate for the NDR when compared to the PDR is 85% for adults with diabetes [10].
2.2. Source and Study Populations
The source population was identified in the PDR and included all individuals with at least one filled prescription of a blood glucose–lowering drug other than insulins (Anatomic Therapeutic Chemical [ATC] code A10B). Filled prescriptions of liraglutide and semaglutide prescribed for obesity treatment were excluded. The overall study population consisted of new users that were identified from 1 July 2006 to 30 September 2022, allowing for a washout period of 1 year from the launch of PDR. The overall study population was then divided into subpopulations according to the length of follow‐up.
2.3. Follow‐Up Time
For each new user, information on antidiabetic medication use was collected from the study population entry date (date of the first filled prescription of a blood glucose–lowering drug other than insulin from 1 July 2006) up to the date of death, emigration, or 30 September 2022 (i.e., administrative end of the prescription data). The analyses included new users who had 1 year or less and 5 years or less of follow‐up from treatment initiation, respectively. With this design, individuals with treatment durations shorter than 1 or 5 years were still included, reducing selection bias. Further, by including individuals whose medication use was up to, but not longer than 1 or 5 years, the design avoided the inclusion of interrupted ongoing trajectories of antidiabetic medication use. This led to two different subpopulations where individuals with a follow‐up of up to 1 year were also included in the follow‐up of up to 5 years.
2.4. Prescription Duration and Treatment Episodes
Treatment episodes were constructed for new users in the subpopulation with up to 1‐ and 5‐year follow‐ups, respectively. Prescriptions of insulins filled after cohort entry were included in the treatment episodes. Durations of all filled prescriptions included in the study were calculated based on the most common recommended daily doses (Table S1). Treatment episodes were constructed at the drug class level considering the following classes: biguanides (ATC A10BA), sulfonylureas (A10BB), dipeptidyl peptidase 4 inhibitors (A10BH), SGLT2i (A10BK), glucagon‐like peptide 1 agonist (A10BJ), insulins (A10A), and other antidiabetic medications including combination drugs (A10BD, A10BF, A10BG, A10BX). Prescriptions from the same class were longitudinally joined during the follow‐up time [3, 11] (Text S1 and Figure S1).
2.5. Proportion of Days Covered (PDC) by any Antidiabetic Medication
Follow‐up times were divided into twelve 30‐day periods for the follow‐up up to 1 year and fifteen 120‐day periods for the follow‐up up to 5 years. The PDC by each antidiabetic medication class in each period was calculated as the ratio between the total number of days covered by each specific class and 30 or 120 days, respectively. Finally, the overall PDC by any antidiabetic medication in each period was calculated as the average of the PDC by each drug class. The longitudinal antidiabetic medication use of each individual in the study population was defined as a trajectory of 12 or 15 PDC by any antidiabetic medication for up to 1‐ and 5‐year follow‐up, with right‐censored periods (due to death, emigration, or administrative end of data) included as missing values.
2.6. K‐Means Algorithm for Clustering of Longitudinal Trajectories
The individuals' longitudinal trajectories of antidiabetic medication use were clustered using the unsupervised clustering method K‐means algorithm for longitudinal data [8, 12, 13]. K‐means was applied to identify clusters of individuals' average PDC by any antidiabetic medication using the R statistical software package “kml” and allowing k‐means to run from two to six clusters 100 times each to identify the best partition [12, 14, 15] (full description in Text S2).
For the main analysis, the algorithm was applied to the whole population of new users of a blood glucose–lowering drug, including the trajectories with missing values due to right‐censored periods for up to 1‐ and 5‐year follow‐up. Once the optimal number of clusters was identified, the individual's probability of cluster membership was extracted to assign individuals to different clusters of adherence measured by the average PDC by any antidiabetic medication. Clusters of adherence were categorized according to their mean trajectory of PDC (“lower” and “higher” for two clusters and “lower,” “intermediate,” and “higher” for three clusters).
2.7. Supplementary Analyses
In addition to the main analysis, K‐means clustering was applied to the subpopulation of patients with a recorded type 2 diabetes diagnosis in NDR at any time during the study period (including the trajectories with missing values due to right censoring). The K‐means algorithm was also applied to the subpopulation of individuals with complete information on PDC by any antidiabetic medication over the full follow‐up of 1 and 5 years (excluding individuals with right‐censored follow‐up). Finally, since the Covid‐19 pandemic could have impacted treatment initiation, the K‐means algorithm was applied to the subpopulation of individuals that initiated treatment with a blood glucose–lowering drug before January 2020 (including the trajectories with missing values due to right censoring).
2.8. Descriptive Analyses
After clustering of individuals in longitudinal trajectories of PDC by any antidiabetic medication, descriptive information for each cluster was provided, including the proportion of individuals in each cluster, mean (standard deviation [SD]) PDC by any antidiabetic medication, sex, age at treatment initiation (mean and SD), proportion of individuals with different numbers of antidiabetic medication drug classes during follow‐up, and insulin use after treatment initiation. Additional information for individuals with a type 2 diabetes diagnosis in NDR was provided, including HbA1c (mean and SD), proportion of individuals with a HbA1c under the target value of 53 mmol/mol as recommended by the American Diabetes Association and the European Association for the Study of Diabetes [16], mean glucose value in the last 2 weeks preceding the visit (mean and SD), body mass index (BMI; mean, SD and weight categories), estimated glomerular filtration rate (eGFR; mean and SD), frequency of physical activity (never, < 1 time per week, regularly 1–2 times per week, regularly 3–5 times per week, daily), and tobacco habits including frequency of smoking and snuffing (daily, never, occasionally, and stopped). Information from NDR was extracted using the visit date closest to the date of cohort entry. Missing values in the continuous variables HbA1c, mean glucose value in the last 2 weeks preceding the visit, BMI, and eGFR were excluded from the calculation of mean and SD.
3. Results
3.1. Clusters of Adherence for New Users of a Blood Glucose–Lowering Drug
The subpopulation with follow‐up of up to 1 year was composed of 75,421 new users who filled a total of 346,156 prescriptions. These new users were also included in the subpopulation of individuals with follow‐up of up to 5 years, which was composed of 283,795 individuals filling 4,099,953 prescriptions.
The subpopulations with up to 1 and 5 years of follow‐up had an overall mean PDC of 0.7 and 0.3, respectively (Table 1). Applied separately in the two subpopulations, K‐means clustering identified as the best partition the one with two adherence groups defined by trajectories of overall PDC by any antidiabetic medication (Figure 1). For the 1‐year follow‐up, 70.6% of individuals was in the cluster with higher mean PDC (0.9) and 29.4% in the cluster with lower mean PDC (0.4), while the corresponding figures for the 5‐year follow‐up were 36.9% (higher mean PDC [0.9]) and 63.1% (lower mean PDC [0.3]), with mean adherence decreasing mainly during the first year.
TABLE 1.
Characteristics of new users of a blood glucose–lowering drug in clusters of adherence measured by the PDC by any antidiabetic medication for up to 1‐ and 5‐year follow‐up.
| Up to 1‐year follow‐up | Up to 5‐year follow‐up | |||||
|---|---|---|---|---|---|---|
| Cluster of adherence | Cluster of adherence | |||||
| Total | Higher | Lower | Total | Higher | Lower | |
| Number of individuals | n = 75,421 | n = 53,223 (70.6%) | n = 22,198 (29.4%) | n = 283,795 | n = 104,710 (36.9%) | n = 179,085 (63.1%) |
| Average PDC | ||||||
| Mean (SD) | 0.7 (0.3) | 0.9 (0.1) | 0.4 (0.1) | 0.3 (0.3) | 0.9 (0.2) | 0.3 (0.1) |
| Sex, n (%) | ||||||
| Men | 42,178 (55.9) | 31,030 (58.3) | 11,148 (50.2) | 155,858 (54.9) | 61,512 (58.7) | 94,346 (52.7) |
| Women | 33,243 (44.1) | 22,193 (41.7) | 11,050 (49.8) | 127,937 (45.1) | 43,198 (41.3) | 84,739 (47.3) |
| Age | ||||||
| Mean (SD) | 65.6 (16.4) | 67.0 (15.5) | 62.1 (18.0) | 63.1 (16.2) | 65.7 (14.8) | 61.5 (16.7) |
| Age group, n (%) | ||||||
| <= 29 | 2252 (3.0) | 1013 (1.9) | 1239 (5.6) | 10,117 (3.6) | 1670 (1.6) | 8447 (4.7) |
| 30–49 | 10,380 (13.8) | 6229 (11.7) | 4151 (18.7) | 45,691 (16.1) | 13,036 (12.4) | 32,655 (18.2) |
| 50–69 | 29,150 (38.6) | 20,909 (39.3) | 8241 (37.1) | 123,140 (43.4) | 46,626 (44.5) | 76,514 (42.7) |
| 70+ | 33,639 (44.6) | 25,072 (47.1) | 8567 (38.6) | 104,847 (36.9) | 43,378 (41.4) | 61,469 (34.3) |
| Number of antidiabetic drug classes used during follow‐up, n (%) | ||||||
| 1 | 61,822 (82.0) | 45,926 (86.3) | 15,896 (71.6) | 188,659 (66.5) | 81,805 (78.1) | 106,854 (59.7) |
| 2 | 11,605 (15.4) | 6259 (11.8) | 5346 (24.1) | 67,837 (23.9) | 18,388 (17.6) | 49,449 (27.6) |
| 3 | 1722 (2.3) | 931 (1.7) | 791 (3.6) | 20,175 (7.1) | 3880 (3.7) | 16,295 (9.1) |
| 4 | 220 (0.3) | 99 (0.2) | 121 (0.5) | 5627 (2.0) | 575 (0.5) | 5052 (2.8) |
| 5 | 41 (0.1) | 8 (0) | 33 (0.1) | 1275 (0.4) | 62 (0.1) | 1213 (0.7) |
| 6 | 8 (0) | 0 (0) | 8 (0) | 200 (0.1) | 0 (0) | 200 (0.1) |
| 7 | 3 (0) | 0 (0) | 3 (0) | 22 (0) | 0 (0) | 22 (0) |
| Insulin use during follow‐up, n (%) | ||||||
| No | 69,822 (92.6) | 50,164 (94.3) | 19,658 (88.6) | 249,065 (87.8) | 96,439 (92.1) | 152,626 (85.2) |
| Yes | 5599 (7.4) | 3059 (5.7) | 2540 (11.4) | 34,730 (12.2) | 8271 (7.9) | 26,459 (14.8) |
Note: Percentages for “Number of Individuals” in the cluster are calculated by row (% in the total population), the percentages for all the other variables are calculated by column (% in the cluster).
Abbreviations: %, percentage; n, number of individuals; PDC, proportion of days covered; SD, standard deviation.
FIGURE 1.
Mean trajectories of PDC by any antidiabetic medication in new users of a blood glucose–lowering drug: (A) Follow‐up up to 1 year; (B) follow‐up up to 5 years. PDC = Proportion of days covered; ADM = antidiabetic medications; (A) Green mean trajectory = 70.6% of the study population with mean PDC = 0.9; Light blue mean trajectory = 29.4% of the study population with mean PDC = 0.4; (B) Green mean trajectory = 63.1% of the study population with mean PDC = 0.3; Light blue mean trajectory = 36.9% of the study population with mean PDC = 0.9.
For the 5‐year follow‐up, when compared to the group with lower mean PDC and to the total population included in the k‐mean clustering, individuals in the group with higher mean PDC had a slightly higher proportion of men (58.7% vs. 52.7% in the lower mean PDC group and 54.9% in the total population) and a slightly lower proportion of women (41.3% vs. 47.3% in the lower mean PDC group and 45.1% in the total population). Additionally, individuals in the group with higher mean PDC were older, with a mean age of 65.7 versus 61.5 in the lower mean PDC group and 63.1 years in the total population (with 41.4% being more than 70 years old at treatment initiation vs. 34.3% and 36.9% in the lower mean PDC group and the total population, respectively). A higher proportion of individuals in the group with higher mean PDC had filled prescriptions for only one antidiabetic drug class (78.1% vs. 59.7% in the lower mean PDC group and 66.5% in the total population) and a lower proportion initiated insulin treatment during follow‐up (7.9% vs. 14.8% in the lower mean PDC group and 12.2% in the total population). Similar patterns of characteristics were observed for the clusters identified in the 1‐year follow‐up (Table 1).
3.2. Clusters of Adherence in Subpopulations With a Type 2 Diabetes Diagnosis, Complete Information on Antidiabetic Medication Use, or Initiating Treatment Prior to the COVID‐19 Pandemic
In the subpopulation of new users with a recorded diagnosis of type 2 diabetes, a total of 30,502 and 188,264 individuals were included in the k‐means clustering of trajectories, with up to 1‐ and 5‐year follow‐up, and with an overall mean PDC of 0.7 and 0.4, respectively (Table 2). As in the main analysis, the maximization of the quality criterion led to a best partition with two adherence groups (Figure 2). For up to 1‐year follow‐up, 67.8% of individuals were in the cluster with higher mean PDC (0.9), while for up to 5‐year follow‐up, only 33.3% were in the cluster with the higher mean PDC (0.9). The clusters with lower mean PDC (0.4 and 0.3 for 1 and 5 years, respectively) included the remaining 32.2% and 66.7% of the populations. Characteristics of individuals in the two clusters followed similar patterns seen in the main analysis, with less pronounced differences in proportions. For both follow‐up periods and when compared to the cluster with lower mean PDC and to the total population included in the k‐means clustering, the clusters with higher mean PDC included a slightly higher proportion of men, of older ages, filling prescriptions from only one antidiabetic drug class, and not initiating insulin during follow‐up. Additionally, when compared to the cluster with lower mean PDC and to the total population, the clusters with higher mean PDC included a slightly lower proportion of individuals with an HbA1c < 53 mmol/mol, with normal BMI, and smoking, but slightly higher proportions of both daily and never snuff users. Negligible differences between the clusters of adherence were observed for the mean eGFR values and weekly physical activity.
TABLE 2.
Characteristics of new users of a blood glucose–lowering drug in clusters of adherence measured by the PDC by any antidiabetic medication for up to 1‐ and 5‐year follow‐up—subpopulation of patients with a recorded type 2 diabetes diagnosis in NDR.
| Up to 1‐year follow‐up | Up to 5‐year follow‐up | |||||
|---|---|---|---|---|---|---|
| Cluster of adherence | Cluster of adherence | |||||
| Total | Higher | Lower | Total | Higher | Lower | |
| Number of individuals | n = 30,502 | n = 20,671 (67.8%) | n = 9831 (32.2%) | n = 188,264 | n = 62,607 (33.3%) | n = 125,657 (66.7%) |
| Average PDC | ||||||
| Mean (SD) | 0.7 (0.3) | 0.9 (0.1) | 0.4 (0.1) | 0.4 (0.3) | 0.9 (0.1) | 0.3 (0.1) |
| Sex, n (%) | ||||||
| Men | 17,749 (58.2) | 12,165 (58.9) | 5584 (56.8) | 110,293 (58.6) | 37,360 (59.7) | 72,933 (58.0) |
| Women | 12,753 (41.8) | 8506 (41.1) | 4247 (43.2) | 77,971 (41.4) | 25,247 (40.3) | 52,724 (42.0) |
| Age | ||||||
| Mean (SD) | 66.8 (14.0) | 67.1 (13.8) | 66.1 (14.5) | 64.6 (13.7) | 65.3 (13.4) | 64.3 (13.9) |
| Age group, n (%) | ||||||
| < 29 | 292 (1.0) | 157 (0.8) | 135 (1.4) | 2005 (1.1) | 428 (0.7) | 1577 (1.3) |
| 30–49 | 3354 (11.0) | 2143 (10.4) | 1211 (12.3) | 24,676 (13.1) | 7357 (11.8) | 17,319 (13.8) |
| 50–69 | 13,421 (44.0) | 9222 (44.6) | 4199 (42.7) | 91,468 (48.6) | 31,021 (49.5) | 60,447 (48.1) |
| 70+ | 13,435 (44.0) | 9149 (44.3) | 4286 (43.6) | 70,115 (37.2) | 23,801 (38) | 46,314 (36.9) |
| Number of antidiabetic drug classes used during follow‐up, n (%) | ||||||
| 1 | 23,151 (75.9) | 16,900 (81.8) | 6251 (63.6) | 114,130 (60.6) | 46,185 (73.8) | 67,945 (54.1) |
| 2 | 6147 (20.2) | 3207 (15.5) | 2940 (29.9) | 50,857 (27.0) | 13,025 (20.8) | 37,832 (30.1) |
| 3 | 1028 (3.4) | 512 (2.5) | 516 (5.2) | 16,968 (9.0) | 2917 (4.7) | 14,051 (11.2) |
| 4 | 131 (0.4) | 49 (0.2) | 82 (0.8) | 4967 (2.6) | 432 (0.7) | 4535 (3.6) |
| 5 | 34 (0.1) | 3 (0) | 31 (0.3) | 1142 (0.6) | 48 (0.1) | 1094 (0.9) |
| 6 | 8 (0) | 0 (0) | 8 (0.1) | 179 (0.1) | 0 (0) | 179 (0.1) |
| 7 | 3 (0) | 0 (0) | 3 (0) | 21 (0) | 0 (0) | 21 (0) |
| Insulin use during follow‐up, n (%) | ||||||
| No | 28,039 (91.9) | 19,317 (93.4) | 8722 (88.7) | 165,080 (87.7) | 57,450 (91.8) | 107,630 (85.7) |
| Yes | 2463 (8.1) | 1354 (6.6) | 1109 (11.3) | 23,184 (12.3) | 5157 (8.2) | 18,027 (14.3) |
| Mean glucose value last 2 weeks preceding the visit (rtCGM/isCGM) | ||||||
| Mean and (SD) | 8.8 (2.3) | 8.9 (2.0) | 8.7 (2.7) | 8.1 (2.4) | 8.6 (2.4) | 7.7 (2.4) |
| HbA1c (mmol/mol) | ||||||
| Mean and (SD) | 56.8 (19.0) | 57.7 (19.8) | 55.0 (17.1) | 57.1 (19.3) | 59.6 (20.9) | 55.9 (18.3) |
| HbA1c target (mmol/mol), n (%) | ||||||
| < 53 | 15,623 (51.2) | 10,268 (49.7) | 5355 (54.5) | 95,927 (51.0) | 29,163 (46.6) | 66,764 (53.1) |
| 53+ | 11,350 (37.2) | 7994 (38.7) | 3356 (34.1) | 74,420 (39.5) | 27,656 (44.2) | 46,764 (37.2) |
| Missing | 3529 (11.6) | 2409 (11.7) | 1120 (11.4) | 17,917 (9.5) | 5788 (9.2) | 12,129 (9.7) |
| BMI | ||||||
| Mean and (SD) | 31.0 (6.3) | 31.2 (6.3) | 30.8 (6.3) | 31.2 (6.1) | 31.3 (6.2) | 31.1 (6.1) |
| BMI categories, n (%) | ||||||
| Normal weight | 2702 (8.9) | 1744 (8.4) | 958 (9.7) | 15,814 (8.4) | 5053 (8.1) | 10,761 (8.6) |
| Obesity | 10,744 (35.2) | 7427 (35.9) | 3317 (33.7) | 70,485 (37.4) | 23,617 (37.7) | 46,868 (37.3) |
| Overweight | 6796 (22.3) | 4532 (21.9) | 2264 (23) | 43,530 (23.1) | 14,236 (22.7) | 29,294 (23.3) |
| Underweight | 102 (0.3) | 65 (0.3) | 37 (0.4) | 416 (0.2) | 131 (0.2) | 285 (0.2) |
| Missing | 10,158 (33.3) | 6903 (33.4) | 3255 (33.1) | 58,019 (30.8) | 19,570 (31.3) | 38,449 (30.6) |
| eGFR (ml/min) | ||||||
| Mean and (SD) | 85.0 (26.0) | 85.0 (26.2) | 85.1 (25.7) | 87.0 (25.9) | 87.4 (26.4) | 86.8 (25.6) |
| Physical activity, n (%) | ||||||
| Daily | 5533 (18.1) | 3814 (18.5) | 1719 (17.5) | 36,256 (19.3) | 12,029 (19.2) | 24,227 (19.3) |
| Less than one time per week | 2562 (8.4) | 1739 (8.4) | 823 (8.4) | 17,360 (9.2) | 5648 (9.0) | 11,712 (9.3) |
| Never | 2916 (9.6) | 1961 (9.5) | 955 (9.7) | 18,279 (9.7) | 6219 (9.9) | 12,060 (9.6) |
| Regularly 1–2 times per week | 3296 (10.8) | 2197 (10.6) | 1099 (11.2) | 21,742 (11.5) | 7029 (11.2) | 14,713 (11.7) |
| Regularly 3–5 times per week | 3795 (12.4) | 2592 (12.5) | 1203 (12.2) | 24,873 (13.2) | 8270 (13.2) | 16,603 (13.2) |
| Missing | 12,400 (40.7) | 8368 (40.5) | 4032 (41.0) | 69,754 (37.1) | 23,412 (37.4) | 46,342 (36.9) |
| Smoking, n (%) | ||||||
| No | 18,219 (59.7) | 12,446 (60.2) | 5773 (58.7) | 114,852 (61.0) | 38,397 (61.3) | 76,455 (60.8) |
| Yes | 3049 (10.0) | 1956 (9.5) | 1093 (11.1) | 21,042 (11.2) | 6589 (10.5) | 14,453 (11.5) |
| Missing | 9234 (30.3) | 6269 (30.3) | 2965 (30.2) | 52,370 (27.8) | 17,621 (28.1) | 34,749 (27.7) |
| Smoking frequency, n (%) | ||||||
| Daily | 2494 (8.2) | 1624 (7.9) | 870 (8.8) | 16,935 (9.0) | 5421 (8.7) | 11,514 (9.2) |
| Never | 10,729 (35.2) | 7328 (35.5) | 3401 (34.6) | 67,521 (35.9) | 22,627 (36.1) | 44,894 (35.7) |
| Occasionally | 231 (0.8) | 160 (0.8) | 71 (0.7) | 1739 (0.9) | 547 (0.9) | 1192 (0.9) |
| Stopped | 6188 (20.3) | 4333 (21.0) | 1855 (18.9) | 37,328 (19.8) | 12,929 (20.7) | 24,399 (19.4) |
| Missing | 10,860 (35.6) | 7226 (35.0) | 3634 (37.0) | 64,741 (34.4) | 21,083 (33.7) | 43,658 (34.7) |
| Snuffing frequency, n (%) | ||||||
| Daily | 1792 (5.9) | 1267 (6.1) | 525 (5.3) | 9650 (5.1) | 3815 (6.1) | 5835 (4.6) |
| Never | 10,067 (33.0) | 6937 (33.6) | 3130 (31.8) | 55,234 (29.3) | 19,941 (31.9) | 35,293 (28.1) |
| Occasionally | 106 (0.3) | 68 (0.3) | 38 (0.4) | 663 (0.4) | 245 (0.4) | 418 (0.3) |
| Stopped | 865 (2.8) | 630 (3.0) | 235 (2.4) | 4468 (2.4) | 1772 (2.8) | 2696 (2.1) |
| Missing | 17,672 (57.9) | 11,769 (56.9) | 5903 (60.0) | 118,249 (62.8) | 36,834 (58.8) | 81,415 (64.8) |
Note: Missing values were excluded from the calculation of mean and SD for the continuous variables HbA1c, mean glucose value in the last 2 weeks preceding the visit, BMI, and eGFR.
Percentages for “Number of Individuals” in the cluster are calculated by row (% in the total population), the percentages for all the other variables are calculated by column (% in the cluster).
Abbreviations: %, percentage; BMI, body mass index; eGFR, glomerular filtration rate; HbA1c, hemoglobin A1c; n, number of individuals; PDC, proportion of days covered; rtCGM/isCGM, Real‐time continuous glucose monitoring/FreeStyle Libre glucose monitoring systems; SD, standard deviation; Snuffing, use of a type of smokeless tobacco.
FIGURE 2.
Mean trajectories of PDC by any antidiabetic medication in new users of a blood glucose–lowering drug with a recorded type 2 diabetes diagnosis in NDR: (A) Follow‐up up to 1 year; (B) follow‐up up to 5 years. PDC = Proportion of days covered; ADM = antidiabetic medications; (A) Green mean trajectory = 67.8% of the study population with mean PDC = 0.9; Light blue mean trajectory = 32.2% of the study population with mean PDC = 0.4; (B) Green mean trajectory = 66.7% of the study population with mean PDC = 0.3; Light blue mean trajectory = 33.3% of the study population with mean PDC = 0.9.
Finally, for the 1‐year follow‐up, results from both the analysis in the subpopulation with complete antidiabetic medication use information and the analysis in the subpopulation initiating treatment before the COVID‐19 pandemic led consistently to two main adherence clusters (Figures S2 and S3 panels A). For the 5‐year follow‐up instead, the analysis in the subpopulation with complete antidiabetic medication use information led to three adherence clusters (Figures S2 and S3; panel B), while the analysis in the subpopulation initiating treatment before the COVID‐19 pandemic led to two adherence clusters. In line with the main analysis, the clusters with higher mean PDC included more men, older individuals, those using only antidiabetic medication from one drug class, and individuals who did not initiate insulin treatment during follow‐up (Tables S2 and S3). A smaller cluster of higher mean PDC (0.9) including 10.7% of individuals was found in the subpopulation with complete antidiabetic medication use information for up to 5 years, while 53.1% and 36.2% of individuals were clustered in the lower mean PDC (0.2) and intermediate mean PDC (0.4) clusters, respectively (Figure S2 panel B).
4. Discussion
4.1. Main Findings
This longitudinal study on adherence to any antidiabetic medication in routine clinical care showed that longitudinal adherence over a 5‐year follow‐up was low. Trajectories of PDC by any antidiabetic medication use over the follow‐up of up to 5 years showed that in the lower adherence group, mean adherence decreased mainly during the first year. Despite the importance of high adherence, only 36.9% of the individuals were found in the cluster with higher adherence. This group included more men, older individuals, users of drugs from only one antidiabetic medication class, and individuals not initiating insulin treatment during follow‐up when compared to the total population included in the k‐means clustering and to individuals in the lower adherence group. Notably, the higher adherence group also had slightly higher mean HbA1c and slightly lower proportions of individuals with BMI within the normal weight range at baseline; further, there was a slightly higher proportion of nonsmokers, but there were slightly higher proportions of both daily and never snuff users. Notwithstanding some differences in characteristics among low and high adherent individuals, the study showed a substantial problem with adherence for all patient groups, which urges interventions.
4.2. Research in Context
Measuring adherence to prescribed medication is challenging, and the majority of studies investigating antidiabetic medication used a cross‐sectional study design to collect data via structured questionnaires [4]. Lack of longitudinal data from routine clinical care on adherence to antidiabetic medication in patients with type 2 diabetes has led researchers to develop in silico trials based on simulated data aiming at investigating the effects of long‐term nonadherence to treatments [17]. This literature gap also drives the use of other sources such as data from questionnaire‐based research or register‐based research. The current study used real‐world observational data from Swedish national health registers to provide evidence to be used in clinical practice.
In line with our findings, a recent cross‐sectional study from Iraq reported that nonadherence was more prevalent among women (60.3%) than men (39.7%) [18]. Likewise, an American study using Medicare claims data showed that being a woman was associated with poor adherence to SGLT2i [5]. Another recent American study showed that women had a higher risk of discontinuing, switching, or intensifying antidiabetic treatment after 1 year from treatment initiation [19]. A review article on sex and gender differences in antidiabetic drugs concluded that women have poorer metabolic control, which could be linked to the mechanism of actions of specific antidiabetic drugs and to the underrepresentation of women in clinical trials performed to test the efficacy and safety of medication [20]. On the other hand, in recent years, the properties of metformin, the most commonly recommended first‐line treatment for type 2 diabetes, have been explored for the potential treatment of polycystic ovary syndrome, particularly in women who undergo in vitro fertilization [21]; this could be partially linked to the higher prevalence of women in the lower adherence group observed in this study.
As seen in the current study, work conducted in Saudi Arabia found that older patients were more likely to adhere to antidiabetic medication [22]. The older age in the group with higher adherence seen in the current study could be linked to the fact that in Sweden nursing homes usually have prepackaged medication delivered at regular intervals, which could improve adherence behaviors, alleviating patients from the burden of collecting medication at the pharmacies.
A German study using insurance data investigated the effect of fixed‐dose (multiple active components in a single dosage form) versus loose‐dose (multiple dosage forms) therapies for the treatment of type 2 diabetes. The study showed that fixed‐dose therapies improved adherence in patients in need of more complex treatment strategies [23]. In line with these findings, the current study showed that a high proportion of patients clustered in the higher adherence group were using drugs from only one drug class, and fewer initiated insulin during follow‐up, suggesting that simplified treatment strategies may be linked to better adherence. Complex treatment strategies composed of multiple blood glucose–lowering drugs, including insulin, may be prescribed to reach a better metabolic control when previous strategies fail, or for economic reasons related to the different costs of antidiabetic drugs. The current study also showed a higher proportion of insulin users in the lower adherence clusters. This could be due to the fact that insulin is recommended for the improvement of blood glucose levels when the use of first‐line treatments for type 2 diabetes does not reach desired results due to lack of effectiveness or poor adherence.
4.3. Limitations
In this study, information on filled prescriptions was used to define a proxy measure for patients' adherence in routine clinical care using the PDC. The measure reflects patients' refill behaviors and is a valid indicator of the availability of the medication at the pharmacies. However, as with any measure of adherence based on prescription fills, the PDC can lead to an underestimation of adherence levels when applied in contexts relying on fragmented data, for example, where the PDC is calculated on data based only on one pharmacy. Nevertheless, in our study, we used the Swedish PDR, which has a complete coverage of all prescriptions filled at Swedish pharmacies [9]. Additionally, the PDC based on prescription fills does not allow establishing with certainty that patients have actually consumed their medication as prescribed, which could lead to an overestimation of the actual adherence levels. Nonetheless, the potential bias introduced in this scenario should not be differential, for example, women and men in the study population should have the same probability of filling a prescription without actually taking the medication. Finally, the method has been advocated by the Pharmacy Quality Alliance as the preferred quality indicator for estimating adherence to therapies for chronic diseases [24].
4.3.1. Contribution to the Field, Clinical Implications, and Future Directions
To the best of our knowledge, this is the first large study that investigated overall longitudinal adherence to any antidiabetic medication from treatment initiation over a follow‐up of up to 5 years using advanced methods to identify treatment episodes of continuous use and cluster individuals. Allowing patients to be users of any antidiabetic drug during follow‐up, the study provided an overall picture of adherence levels independent from the type of drug prescribed, thereby increasing generalizability to patients undergoing any treatment regimen. Hence, it expands on prior studies focused on individual drug classes and adherence only during the first year [5, 6]. Additionally, this study investigated the subpopulations of patients with a confirmed diagnosis of type 2 diabetes and additional information collected in clinical practice such as HbA1c and BMI, with complete information on medication use during the 5‐year follow‐up, or including only new users identified prior to the COVID‐19 pandemic. Measurement of adherence is a challenging field of pharmacoepidemiology, and in this study, drug duration was defined at the drug substance level with treatment episodes longitudinally joined at the drug class level. This allowed consideration of the adherence measure switching of drugs within the same class and at the same time to include the time of exposure to other antidiabetic drug classes. Clustering of longitudinal trajectories of adherence over long follow‐up periods is also computationally intense, and this study may be the first thorough attempt to disentangle the complex issue of adherence using large cohorts of new users of antidiabetic medication.
Some differences in characteristics between adherence groups observed in this study may give insights on which patients could be at risk of low adherence. For example, the choice of treatment regimens mattered. While a complex treatment strategy with more than one medication may be needed, the risk of lower adherence must be considered. Most importantly, the study showed a substantial problem with adherence overall, and the knowledge gained from this study could be used, for instance, to inform intervention strategies and future research disentangling heterogeneities in treatment effects in type 2 diabetes. Such interventions may include improved education in disease management for patients, with a special focus on adherence to treatment, as well as health care practitioner training focused on improving treatment choices aiming at sustaining adherence, with the ultimate goal of reducing long‐term diabetes complications. For example, follow‐up strategies targeted at assessing adherence could help clinicians identify latent problems related to treatment effectiveness and improve continuous treatment management. Finally, new technologies such as mobile‐app‐based interventions have shown potential to improve mean HbA1c levels via enhancement of medication adherence in patients with T2D [24]. The adoption of these new technologies in routine clinical care, together with educational interventions, could help achieve proper treatment control in T2D. The findings showed that adherence decreased mostly during the first year, and this result may guide health care practitioners to be cautious about low adherence during this time. Potentially, a closer monitoring of individuals at risk of low adherence during the initial treatment phase could deepen the understanding of the underlying reasons and mitigate those issues to reduce the risk of stopping medication.
Author Contributions
Laura Pazzagli, Björn Pasternak, and Ingvild Odsbu: designed the research. Laura Pazzagli: wrote the manuscript. Laura Pazzagli, Björn Pasternak, Ingvild Odsbu, Carolyn E. Cesta, Ylva Trolle Lagerros, and Rino Bellocco: performed the research. Laura Pazzagli: analyzed the data. Laura Pazzagli 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.
Ethics Statement
The study was approved by the Swedish Ethical Review Authority (dnr 2021–03957, 2023–03824‐02). Informed consent to be included in national registers is not required.
Conflicts of Interest
C. E. C. and I. O. report participation in research projects funded by pharmaceutical companies, all regulator‐mandated phase IV studies, with all funds paid to their institutions (no personal fees) and no relation to the work reported in this paper. All other authors declared no competing interest in this work.
Supporting information
Data S1.
Funding: This study and LP were supported by a grant from FORTE Swedish Research Council for Health, Working Life, and Welfare (project no. 2021–01080). The study is also supported by the Strategic Research Programme (SRP) in Diabetes at Karolinska Institutet (LP) and Karolinska Institutet foundation and funds (LP). BP was supported by a consolidator investigator grant from Karolinska Institutet. The funders had no role in the study design, data collection, interpretation, and statistical analyses. Neither did they influence the decision to publish or the preparation of the manuscript.
Data Availability Statement
Personal data used in this study are protected by privacy laws in Sweden, and therefore, are not publicly available.
References
- 1. Committee A. D. A. P. P., “Pharmacologic Approaches to Glycemic Treatment: Standards of Care in Diabetes—2024,” Diabetes Care 47, no. Supplement_1 (2023): S158–S178. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2. Lambert J., Chekroun M., Gilet H., Acquadro C., and Arnould B., “Assessing Patients' Acceptance of Their Medication to Reveal Unmet Needs: Results From a Large Multi‐Diseases Study Using a Patient Online Community,” Health and Quality of Life Outcomes 16, no. 1 (2018): 134. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3. Pazzagli L., Brandt L., Linder M., et al., “Methods for Constructing Treatment Episodes and Impact on Exposure‐Outcome Associations,” European Journal of Clinical Pharmacology 76, no. 2 (2020): 267–275. [DOI] [PubMed] [Google Scholar]
- 4. Studer C. M., Linder M., and Pazzagli L., “A Global Systematic Overview of Socioeconomic Factors Associated With Antidiabetic Medication Adherence in Individuals With Type 2 Diabetes,” Journal of Health, Population, and Nutrition 42, no. 1 (2023): 122. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5. Hawley C. E., Lauffenburger J. C., Paik J. M., Wexler D. J., Kim S. C., and Patorno E., “Three Sides to the Story: Adherence Trajectories During the First Year of SGLT2 Inhibitor Therapy Among Medicare Beneficiaries,” Diabetes Care 45, no. 3 (2022): 604–613. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6. Kim T., Campbell D. B., Au F., Ronksley P. E., Weaver R. G., and Campbell D. J. T., “Patterns of Metformin Use and Glycated Haemoglobin Trends Among Patients With Newly Diagnosed Type 2 Diabetes in Alberta, Canada,” Diabetes, Obesity & Metabolism 26, no. 4 (2024): 1234–1243. [DOI] [PubMed] [Google Scholar]
- 7. Chen Y. L., Nguyen P. A., Chien C. H., Hsu M. H., Liou D. M., and Yang H. C., “Machine Learning‐Based Prediction of Medication Refill Adherence Among First‐Time Insulin Users With Type 2 Diabetes,” Diabetes Research and Clinical Practice 207 (2024): 111033. [DOI] [PubMed] [Google Scholar]
- 8. Genolini C. and Falissard B., “KmL: K‐Means for Longitudinal Data,” Computational Statistics 25, no. 2 (2010): 317–328. [Google Scholar]
- 9. Wettermark B., Hammar N., Fored C. M., et al., “The New Swedish Prescribed Drug Register–Opportunities for Pharmacoepidemiological Research and Experience From the First Six Months,” Pharmacoepidemiology and Drug Safety 16, no. 7 (2007): 726–735. [DOI] [PubMed] [Google Scholar]
- 10. Eeg‐Olofsson K., Åkesson K., Thorén A., et al., “Annual Report 2023. National Diabetes Register,” 2024.
- 11. Pazzagli L., Andersen M., and Sessa M., “Pharmacological and Epidemiological Considerations While Constructing Treatment Episodes Using Observational Data: A Simulation Study,” Pharmacoepidemiology and Drug Safety 31, no. 1 (2022): 55–60. [DOI] [PubMed] [Google Scholar]
- 12. Genolini C. and Falissard B., “Kml: A Package to Cluster Longitudinal Data,” Computer Methods and Programs in Biomedicine 104, no. 3 (2011): e112–e121. [DOI] [PubMed] [Google Scholar]
- 13. Genolini C., Alacoque X., Sentenac M., and Arnaud C., “Kml and kml3d: R Packages to Cluster Longitudinal Data,” Journal of Statistical Software 65, no. 4 (2015): 1–34. [Google Scholar]
- 14. Jain A. K., “Data Clustering: 50 Years Beyond K‐Means,” Pattern Recognition Letters 31, no. 8 (2010): 651–666. [Google Scholar]
- 15. Caliński T. and Harabasz J., “A Dendrite Method for Cluster Analysis,” Communications in Statistics 3, no. 1 (1974): 1–27. [Google Scholar]
- 16. Davies M. J., Aroda V. R., Collins B. S., et al., “Management of Hyperglycemia in Type 2 Diabetes, 2022. A Consensus Report by the American Diabetes Association (ADA) and the European Association for the Study of Diabetes (EASD),” Diabetes Care 45, no. 11 (2022): 2753–2786. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17. Visentin R., Cobelli C., Sieber J., and Dalla Man C., “Short‐ and Long‐Term Effects on Glucose Control of Nonadherence to Insulin Therapy in People With Type 2 Diabetes an in Silico Study,” Journal of Diabetes Science and Technology 18, no. 2 (2024): 309–317. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18. Hamalaw S., Hama Salih A., and Weli S., “Non‐Adherence to Anti‐Diabetic Prescriptions Among Type 2 Diabetes Mellitus Patients in the Kurdistan Region of Iraq,” Cureus 16, no. 5 (2024): e60572. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19. Liss D. T., Cherupally M., O'Brien M. J., et al., “Treatment Modification After Initiating Second‐Line Medication for Type 2 Diabetes,” American Journal of Managed Care 29, no. 12 (2023): 661–668. [DOI] [PubMed] [Google Scholar]
- 20. Campesi I., Seghieri G., and Franconi F., “Type 2 Diabetic Women Are Not Small Type 2 Diabetic Men: Sex‐and‐Gender Differences in Antidiabetic Drugs,” Current Opinion in Pharmacology 60 (2021): 40–45. [DOI] [PubMed] [Google Scholar]
- 21. Johnson N. P., “Metformin Use in Women With Polycystic Ovary Syndrome,” Annals of Translational Medicine 2, no. 6 (2014): 56. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22. Khardali A., Kashan Syed N., Alqahtani S. S., et al., “Assessing Medication Adherence and Their Associated Factors Amongst Type‐2 Diabetes Mellitus Patients of Jazan Province, Saudi Arabia: A Single‐Center, Cross‐Sectional Study,” Saudi Pharmaceutical Journal 32, no. 1 (2024): 101896. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23. Böhm A.‐K., Schneider U., Aberle J., and Stargardt T., “Regimen Simplification and Medication Adherence: Fixed‐Dose Versus Loose‐Dose Combination Therapy for Type 2 Diabetes,” PLoS One 16, no. 5 (2021): e0250993. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24. Shrivastava T. P., Goswami S., Gupta R., and Goyal R. K., “Mobile App Interventions to Improve Medication Adherence Among Type 2 Diabetes Mellitus Patients: A Systematic Review of Clinical Trials,” Journal of Diabetes Science and Technology 17, no. 2 (2023): 458–466. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data S1.
Data Availability Statement
Personal data used in this study are protected by privacy laws in Sweden, and therefore, are not publicly available.