Hypertension affects a significant portion of the global population, yet achieving adequate blood pressure (BP) control remains a challenge, often leading to increased cardiovascular morbidity and mortality. Current treatment guidelines are broad, and the selection of antihypertensive agents typically follows a stepwise, trial-and-error approach that can be slow and suboptimal. A personalized approach, leveraging AI to analyze rich patient data and predict optimal therapy, could dramatically improve treatment efficacy, accelerate BP control, and minimize adverse effects, thereby transforming hypertension management. This work addresses a critical area of chronic disease care.
The aim of the study.
To develop and validate a machine learning-driven clinical decision support (CDS) system capable of providing personalized recommendations for optimal antihypertensive monotherapy or combination therapy based on individual patient phenotypes extracted from electronic health records, with the goal of improving blood pressure control rates.
Material and methods.
This retrospective cohort study utilized comprehensive electronic health record (EHR) data from 80,000 hypertensive adult patients who initiated or switched antihypertensive therapy over a five-year period. The dataset included detailed information on patient demographics, medical history (e.g., diabetes, renal impairment, CAD), baseline and longitudinal laboratory values (e.g., electrolytes, renal function), pharmacogenomic markers (SNPs related to drug metabolism), lifestyle factors (BMI, smoking status), and previous responses to specific antihypertensive medications. The primary outcome was the achievement of sustained blood pressure control (<140/90 mmHg) within three months of initiating a new regimen, coupled with a low incidence of reported side effects. A multi-label XGBoost classifier was trained on 70% of the data to recommend one or more of six major classes of antihypertensive drugs (thiazide diuretics, ACE inhibitors, ARBs, beta-blockers, calcium channel blockers, alpha-blockers), or common combinations, to achieve the outcome. The remaining 30% of the data was used for independent validation. The CDS system's recommendations were compared against the outcomes of physician-selected 'first-line' therapies using accuracy, precision, recall, and F1-score. Explainable AI (XAI) techniques (SHAP values) were used to identify key predictive features.
The machine learning-driven CDS system achieved a high accuracy of 88.5% (95% CI: 87.8-89.2) in recommending an individualized antihypertensive regimen that led to successful blood pressure control. This significantly outperformed the observed effectiveness of empirical physician-selected 'first-line' therapies, which achieved BP control in 72% of cases (p<0.001). The system's precision for recommending a successful therapy was 0.89, and recall was 0.86. Feature importance analysis revealed that interactions between specific genetic markers (e.g., CYP2D6 variants affecting beta-blocker metabolism) and comorbidity clusters (e.g., diabetes with concurrent renal impairment) were highly influential in guiding optimal drug selection. For example, patients with a specific CYP2D6 variant alongside diabetes showed a 2.5-fold higher probability of BP control with a calcium channel blocker-based regimen compared to a beta-blocker. The system provided a 'top-3' list of recommended drugs with associated probabilities of success, with the top recommendation having an average 92% chance of achieving BP control in target patients.
This study demonstrates that an AI-powered Clinical Decision Support system can provide highly accurate and personalized recommendations for antihypertensive therapy, significantly improving the rate of blood pressure control compared to traditional guidelines and empirical approaches. By leveraging rich patient phenotype data from EHRs, including pharmacogenomics and comorbidity patterns, this system offers a crucial step towards precision medicine in hypertension. This promises to reduce therapeutic inertia, minimize adverse events, and ultimately lower the burden of cardiovascular disease. Future work involves prospective clinical trials to evaluate the system's impact on long-term patient outcomes and seamless integration into routine clinical practice.
1. Whelton PK, et al. 2017 ACC/AHA/AAPA/ABC/ACPM/AGS/APhA/ASH/ASPC/NMA/PCNA Guideline for the Prevention, Detection, Evaluation, and Management of High Blood Pressure in Adults. J Am Coll Cardiol. 2018;71(19):e127-e248.
2. Beam AL, Kohane IS. Big Data and Machine Learning in Health Care. JAMA. 2018;319(13):1317-1318.
3. Zhang J, et al. Machine learning for personalized management of hypertension. J Hypertens. 2020;38(11):2100-2109.
4. Shah NR, et al. Machine Learning Models for Clinical Decision Support in Hypertension. Circulation. 2021;144(Suppl 1):A14227.
5. Topol EJ. Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again. Basic Books; 2019.