This retrospective register-based cohort evaluated statin adherence in patients with hypertension who initiated statin treatment between 2012 and 2015 in a large Swedish region. Data from all primary health care centres in the region were linked to registries containing comorbidities, dispensed medications, and socioeconomic variables. The study population comprised 30,497 patients with a mean age of 66 years; 48% were women.
Adherence was quantified over the 2 years following statin initiation using the proportion of days covered (PDC). Adequate adherence was defined as PDC > 80%, consistent with commonly used thresholds for long-term medication adherence assessment.
The mean 2-year PDC in the cohort was 73.2%. Overall, 60.2% of patients achieved adequate adherence (PDC > 80%). Adherence differed significantly by indication for statin therapy (P < 0.001): 54.0% achieved adequate adherence in primary prevention, 61.3% in patients with diabetes but without established cardiovascular disease, and 65.5% in secondary prevention. These results indicate variation in long-term statin use according to underlying cardiovascular risk or indication.
To identify baseline characteristics associated with adherence, the investigators applied an explainable machine learning approach. They trained an XGBoost model and used SHAP values to determine the relative importance of baseline features in predicting whether patients achieved adequate adherence. The use of explainable machine learning allowed ranking of predictors while providing interpretable contributions of individual variables to the model’s predictions.
Findings from the XGBoost/SHAP analysis were complemented by conventional adjusted multilevel regression to quantify associations and to account for clustering by primary health care centre.
According to SHAP-based importance ranking, the five most influential baseline characteristics associated with achieving adequate adherence were:
Dispensed antithrombotic therapy. In adjusted analyses this characteristic was associated with higher odds of adequate adherence (OR 1.38, 95% CI 1.26–1.50).
The patient’s primary health care centre, reflecting a provider-level influence; clustering by centre produced a median odds ratio (MOR) of 1.16 (95% CI 1.12–1.20) indicating measurable between-centre variation in adherence.
Older age: each 10-year increase in age was associated with higher odds of adequate adherence (OR 1.14, 95% CI 1.12–1.17).
Initiation with atorvastatin was associated with greater likelihood of adequate adherence compared with other statins (OR 1.33, 95% CI 1.24–1.44).
Countries of birth outside Sweden were associated with lower odds of adequate adherence; the reported OR range was 0.73–0.78.
These characteristics represent a mixture of clinical treatment factors, sociodemographic attributes, and health-care provider influences.
The machine learning importance ranking was followed by adjusted multilevel regression models that provided effect estimates for associations between baseline characteristics and adequate adherence. The analysis reported odds ratios (ORs) with 95% confidence intervals and quantified between-centre heterogeneity using the median odds ratio (MOR).
Examples of reported effect sizes include OR 1.38 for dispensed antithrombotic therapy (95% CI 1.26–1.50), OR 1.14 per 10-year increase in age (95% CI 1.12–1.17), and OR 1.33 for initiation with atorvastatin (95% CI 1.24–1.44). The MOR for primary health care centre was 1.16 (95% CI 1.12–1.20), illustrating a nontrivial provider-level contribution to adherence variability.
Only about 60% of patients with hypertension initiating statin therapy achieved adequate 2-year adherence, and adherence differed by indication, with the lowest proportion in primary prevention. The combination of patient-level (age, country of birth, concurrent antithrombotic therapy, initial statin choice) and provider-level (primary health care centre) predictors suggests opportunities for targeted interventions.
Provider-level variability implies that interventions at the primary care centre level—such as adherence support, prescription practices, or follow-up systems—could influence long-term statin use. Patient-level factors may guide tailored counseling or support for groups at higher risk of nonadherence.
Strengths of the study include a large, region-wide register-based sample drawn from all primary health care centres in the region and linkage to medication dispensation, comorbidity, and socioeconomic data. The use of explainable machine learning (XGBoost with SHAP) plus conventional multilevel regression combined interpretable variable ranking with adjusted effect estimation.
Limitations inherent to register-based observational research include potential unmeasured confounding and that dispensed medication (PDC) is an indirect measure of actual medication-taking. Specific details on some covariates or analytic choices beyond those reported in the abstract were not provided in the source text.
Conflict of interest: one author reported receiving speaker honoraria from AstraZeneca and Boehringer-Ingelheim for lectures not related to this manuscript.