Statin adherence among patients with hypertension in Swedish primary health care: a QregPV cohort study using explainable machine learning

Figure 1. Study flowchart. Among patients classified as receiving secondary preventive therapy, 1657 were classified based solely on secondary prevention diagnoses, 4664 based solely on secondary prevention medications, and 6967 had both a secondary prevention diagnosis and medication.

Figure 2. SHAP values for the full study cohort (test data) illustrating the 10 baseline characteristics most strongly associated with predicted adherence, defined as proportion of days covered (PDC) > 80%. Each dot represents one individual. The x-axis shows how much each variable contributed to the prediction of adherence in the machine learning model. Positive SHAP values indicate higher predicted adherence, whereas negative values indicate lower predicted adherence. The colours indicate whether the value of the variable was high (red) or low (blue). ACE: angiotensin-converting enzyme.

Figure 3. Forest plot showing effect estimates from multilevel regression models in the full study cohort for the 10 baseline characteristics most strongly associated with adherence, defined as proportion of days covered (PDC) > 80%. Odds ratios and, for primary health care centres, the median odds ratios (MORs) are presented from unadjusted and adjusted analyses, with confounder selection based on the modified disjunctive cause criterion.

Figure 4. SHAP values for the primary prevention group (test data) illustrating the 10 baseline characteristics most strongly associated with predicted adherence, defined as proportion of days covered (PDC) > 80%. Each dot represents one individual. The x-axis shows how much each variable contributed to the prediction of adherence in the machine learning model. Positive SHAP values indicate higher predicted adherence, whereas negative values indicate lower predicted adherence. The colours indicate whether the value of the variable was high (red) or low (blue). ACE: angiotensin-converting enzyme. LDL: low-density lipoprotein.

Figure 5. Forest plot showing effect estimates from multilevel regression models in the primary prevention group for the 10 baseline characteristics most strongly associated with adherence, defined as proportion of days covered (PDC) > 80%. Odds ratios and, for primary health care centres, the median odds ratios (MOR)s are presented from unadjusted and adjusted analyses, with confounder selection based on the modified disjunctive cause criterion.