Prediction of Incident Hypertension Within the Next Year: Prospective Study Using Statewide Electronic Health Records and Machine Learning

Figure 1. Workflow diagram depicting model construction and evaluation. The retrospective cohort consisted of 823,627 individuals with electronic health record (EHR) profiles extracted from 2013, 92,512 of whom (cases) developed hypertension in 2014. The validation cohort consisted of 680,810 individuals, with EHR profiles extracted from 2014, 60,065 of whom received a new hypertension diagnosis in 2015.

Figure 2. (a) The receiver operating characteristic (ROC) curve derived from the prospective cohort (left). (b) Survival curves (right) of the five risk categories identified by the 1-year hypertension risk prediction model. Five risk categories were defined: very low (risk score<0.05), low (risk score 0.05-0.1), medium (risk score 0.1-0.2), high (risk score 0.2-0.4), and very high (risk score >0.4). HR: hazard ratio.

Figure 3. Six classifications of the 169 electronic health record (EHR)–based impactful features recognized by our risk model. COPD: chronic obstructive pulmonary disease; CKD: chronic kidney disease.

Figure 4. Constituent ratios of age and gender (female) subgroups across the identified five risk categories. Age groups (years): <35, 35-49, 50-64, and ≥65.

Figure 5. Kaplan-Meier curves depicting time-to-hypertension for the next 1 year according to coexisting disease subgroup for the very low risk (left) and very high risk (right) categories of the prospective cohort. Disease subgroups comprised patients who received diagnoses of either prehypertension, type 2 diabetes, lipid disorders, mental illness, chronic kidney disease (CKD), or cardiovascular disease (CVD), respectively.

Figure 6. Predicted time-to-hypertension curves for patients (a) who received no diagnosis of other chronic diseases except possible depression and (b) who had at least one diagnosis of other chronic disease. Curves for both subgroups were stratified by age (<65 vs ≥65 years), gender (male vs female), and depression diagnosis (depression vs control).

Figure 7. Patients’ average clinical costs in the past 12 months against the average number of chronic diseases. The balls were formed by 16 common disease subgroups under the very low risk (blue balls) and very high risk (yellow balls) categories, respectively. The ball size indicates the proportion of the disease subgroup under this risk category. The 16 chronic diseases were prehypertension, cardiovascular disease (CVD), type 2 diabetes, lipid disorders, chronic kidney disease (CKD), chronic nephritis, depression, anxiety, schizophrenia, chronic obstructive pulmonary disease (COPD), liver disorders, cancer, anemia, hypokalemia, idiopathic hypersomnia, and hyposmolality or hyponatremia. Reference groups consisted of patients with no diagnosis of any above chronic diseases.

Figure 8. Spearman rank correlation between 12 socioeconomic features and prospective hypertension risk scores. (a) Percentage of people in a county with low income and living more than 1 mile from a supermarket or large grocery store if an urban area, or more than 10 miles from a supermarket or large grocery store if a rural area. (b) The number of supermarkets and grocery stores in the county per 1000 county residents. (c) Percentage of farmers’ markets in the county that sell fresh fruits and vegetables.