Development and Validation of Predictive Models for Non-Adherence to Antihypertensive Medication

Area of Science:

• Cardiovascular Medicine
• Health Informatics
• Artificial Intelligence in Healthcare

Background:

• Suboptimal adherence to antihypertensive medication impacts therapeutic outcomes.
• Artificial intelligence (AI) presents advanced capabilities for analyzing medication adherence data.
• Targeted interventions require accurate identification of patients with low adherence.

Purpose of the Study:

• To develop and validate predictive models for antihypertensive medication non-adherence.
• To utilize patient-reported data and machine learning for adherence prediction.
• To identify key factors associated with medication adherence in hypertensive patients.

Main Methods:

• A cross-sectional study involving 3095 hypertensive patients from community pharmacies.
• Data collection via a structured questionnaire on sociodemographic, medical, and behavioral factors.
• Development and validation of five machine learning models (Logistic Regression, Random Forest, CatBoost, LightGBM, XGBoost) to predict non-adherence using the Adherence to Refills and Medications Scale (ARMS).

Main Results:

• A high prevalence of suboptimal adherence was observed (79.13% with ARMS Score ≥ 15).
• Factors associated with better adherence included frequent blood pressure self-monitoring, reduced salt intake, and pharmacist information.
• The CatBoost model demonstrated the highest performance among the evaluated machine learning models, with ROC AUC scores not exceeding 0.75.

Conclusions:

• Machine learning models were successfully developed and validated for estimating medication non-adherence levels.
• AI shows potential for identifying and stratifying patients based on adherence profiles, despite moderate model performance.
• This study pioneers the use of permutation and SHapley Additive exPlanations feature importance with probability-based adherence stratification for predictive modeling.