Hypertension is highly prevalent in population, while comorbidity and concomitant medications can result in individual treatment risk and benefit. Evidence from clinical trials comes from limited, “purified” population and therefore can be poorly translated into real-world setting. The main objective of this study was to create a model for prediction of blood pressure response for beta-blockers depending on individual patients characteristics based on a real-world clinical data of hypertensive patients.
For real-world efficacy data we used 89000 electronic medical records of patients referred to specialized cardiology clinic due to uncontrolled HTN during the period of January 2010 - December 2016. We extracted data on established risk factors, target organ damage and concomitant metabolic (diabetes, obesity) and cardiovascular diseases (coronary artery disease, heart failure) and used them as individual characteristics that may contribute to treatment response. Antihypertensive efficacy of beta-blockers was determined as more than 10 mmHg decrease of blood pressure level on follow-up visit (within 1–3 months from initial visit with prescription, standard dose, stable doses of all other antihypertensives). We adapted the CART (Classification and Regression Tree) algorithm for building effective personalized antihypertensive treatment rules decision tree. We evaluated model predictive power on the validation set using sensitivity and the specificity.
According to the model, the most significant patient characteristics that affect the effectiveness of beta-blocker therapy were the level of systolic blood pressure (Gini index = 0.36), age (Gini index = 0.29), body mass index (Gini index = 0.18), family history (Gini index = 0.05), dyslipidemia (index Gini = 0.08), chronic heart failure (Gini index = 0.04). However, the model was characterized by high specificity (0.75) with low sensitivity (0.62). This means a strong ability of constructed decision trees to identify patients with ineffective treatment, however, obtained low sensitivity values mean failure to predict treatment effectiveness.
Our initial model confirmed that individual characteristics may predict treatment ineffectiveness of beta-blockers (lower blood pressure, age over 65 and obesity). However improvement of sensitivity of the model is needed for definitive conclusions.
1Almazov National Medical Research Centre - Hypertension Department, Saint-Petersburg, Russia
2ITMO Univerity - Department of High Perfomance Computing, Saint-Petersburg, Russia