Abstract
Background and Objectives: Investigating the adherence to antihypertensive medication and identifying patients with low adherence allows targeted interventions to improve therapeutic outcomes. Artificial intelligence (AI) offers advanced tools for analyzing medication adherence data. This study aimed to develop and validate several predictive models for non-adherence, using patient-reported data collected via a structured questionnaire. Materials and Methods: A cross-sectional, multi-center study was conducted on 3095 hypertensive patients from community pharmacies. A structured questionnaire was administered, collecting data on sociodemographic factors, medical history, self-monitoring behaviors, and informational exposure, alongside medication adherence measured using the Romanian-translated and validated ARMS (Adherence to Refills and Medications Scale). Five machine learning models were developed to predict non-adherence, defined by ARMS quartile-based thresholds. The models included Logistic Regression, Random Forest, and boosting algorithms (CatBoost, LightGBM, and XGBoost). Models were evaluated based on their ability to stratify patients according to adherence risk. Results: A total of 79.13% of respondents had an ARMS Score ≥ 15, indicating a high prevalence of suboptimal adherence. Better adherence was statistically associated (adjusted for age and sex) with more frequent blood pressure self-monitoring, a reduced salt intake, fewer daily supplements, more frequent reading of medication leaflets, and the receipt of specific information from pharmacists. Among the ML models, CatBoost achieved the highest ROC AUC Scores across the non-adherence classifications, although none exceeded 0.75. Conclusions: Several machine learning models were developed and validated to estimate levels of medication non-adherence. While the performance was moderate, the results demonstrate the potential of AI in identifying and stratifying patients by adherence profiles. Notably, to our knowledge, this study represents the first application of permutation and SHapley Additive exPlanations feature importance in combination with probability-based adherence stratification, offering a novel framework for predictive adherence modelling.