Abstract 4370092: Utilizing Machine Learning to Predict Uncontrolled Blood Pressure Status Among Adults with Hypertension

Introduction: Antihypertensive medication is effective at lowering blood pressure (BP). However, clinical inertia (i.e., failure of providers to increase the frequency or dose of antihypertensive medication) is common even when treatment intensification is indicated. Methods: We developed and validated a machine learning algorithm (MLA) to predict uncontrolled BP among adults with hypertension (HTN). Data were obtained from the electronic health record (EHR) of an academic medical center from 2012–2020. Inclusion criteria were >20 years old, diagnosis of hypertension, use of antihypertensive medication, and a primary care visit where the patient had uncontrolled BP (i.e., index visit). HTN was defined as uncontrolled if BP >=140/90 mmHg from 2012-2017 and as BP>=130/80 from 2017-2020. Data were split into training (80%) and testing (20%). The MLA prediction outcome was uncontrolled BP (yes/no). Only structured data from the EHR were used (i.e., demographics, encounters, diagnoses, lab values). We also feature engineered variables, such as BMI, days between encounters, and number of past uncontrolled BP readings over time prior to a visit. Finally, the EHR data were linked to social determinants of health data by zip code tabulation areas. We utilized XGboost for the MLA and included data from the index visit and three prior office visits to predict BP control at the follow up visit. Results: There were 67,603 unique patients and 1,134,644 total primary care encounters. Mean age of patients was 64.5 (SD=12.8) years. Most visits were with women (59.4%) and black (39.0%) and white (35.8%) patients. Following the uncontrolled BP index visit, 53.0% of participants had uncontrolled BP at the follow up visit. Index visit SBP and DBP were 135.0 (SD=17.2) and 78.9 (SD=10.1) mmHg and follow up visit SBP and DBP for patients with persistent uncontrolled BP were 139.4 (SD=18.1) and 80.6 (SD=10.7) mmHg and for controlled BP were 130.1 (SD=14.6) and 77.1 (SD=9.0) mm Hg. The AUC for the XGboost model was 0.792. The accuracy, precision, recall, f-measure, sensitivity, specificity, and positive predictive value for the model was 0.71, 0.67, 0.80, 0.73, 0.80, 0.62, and 0.67, respectively. Conclusions: Using EHR data, an MLA accurately predicted uncontrolled BP status at a primary care visit following a visit where there was an uncontrolled BP reading. This information could potentially be used to target patients for interventions and lower clinical inertia.