Hypertension in this study was defined by the International Classification of Diseases, 9th Revision, Clinical Modification (ICD-9-CM) diagnosis codes from category 401, which refers specifically to essential (or primary) hypertension. For the prediction modeling cohort, cases in the retrospective population referred to patients who received a new diagnosis of essential hypertension during the calendar year of 2014 (from January 1, 2014 to December 31, 2014), whereas cases in the prospective cohort were patients receiving an incident diagnosis of essential hypertension during the 2015 calendar year (from January 1, 2015 to December 31, 2015). Along with the implementation of ICD-10-CM since October 1, 2015, an ICD-10-CM category of I10 was used to define the diagnosis of primary hypertension, which is equivalent to the ICD-9-CM category 401 according to the ICD-10-CM General Equivalence Mapping (GEM), a crosswalk between the two code standards maintained by the Center for Medicare Services and the Centers for Disease Control and Prevention [30].

Given that our prediction model was constructed to predict an individual’s hypertension risk during the following 1 year based on his or her medical records from the current year, feature profiles for the retrospective cohort were extracted from the clinical and health historical record of 2013, giving a total of 823,627 patients, 92,512 of whom developed hypertension in the year 2014.

Samples were excluded from the retrospective cohort if there was any record of a hypertension diagnosis before January 1, 2014. Similarly, the prospective cohort was extracted from the clinical and health historical record of 2014, offering a total of 680,810 samples and 60,065 cases with incident hypertension in the year 2015. Samples were excluded in prospective cohort if there was any record of a hypertension diagnosis before January 1, 2015.

Various categories of data were extracted from the original health records, including demographics, laboratory and radiographic test results, essential and secondary diagnoses and procedures, outpatient medication prescriptions, clinical utility records, as well as a number of accessible socioeconomic variables extracted from the US census and United States Department of Agriculture (USDA) websites [31,32]. Overall, more than 15,000 features were recruited in our original data pool. Among them, laboratory and radiographic test results were coded by logical observation identifier names and codes, outpatient medication prescriptions were coded by the National Drug Code, and primary and secondary diagnoses and procedures were coded by ICD-9-CM. All ICD-10-CM codes were mapped to ICD-9-CM using the GEM tool to keep the data consistent. Some socioeconomic features are directly downloaded from the US census website using the advanced search tool under the American Fact Finder category, and others were derived from the USDA website using the Web-based mapping tool, Food Environment Atlas (see Multimedia Appendix 1). Using zone improvement plan (ZIP) code or county or tract, all these socioeconomic variables are integrated into our knowledge base.

Before feature selection, the k-nearest neighbors (KNN) approach was used to impute missing data. For a given patient with missing values, the KNN method identified the k-nearest patients based on Euclidean distance. Using these patients, missing values were then replaced using a majority vote for discrete variables and weighted means for continuous features. One advantage of using this method is that missing values in all features are imputed simultaneously without the need to treat features individually [33].

Before the machine-learning algorithm, a univariate correlation filtering step was introduced to remove all features that are not directly related to the outcome variable [34]. In general, the Cochran-Mantel-Haenszel test is applied for binary features by investigating the association between the feature and outcome under age-group strata using the following formula: R=∑^K_i=1A_iD_i/T_i/ ∑^K_i=1B_iC_i/T_i, where A_i, B_i, C_i, and D_i are the counts of individuals in 2x2 contingency table of the i- th age group. In addition, the Cochran-Armitage trend test is used to test ordinal variables, and univariate logistic regression is used for continuous variables. After applying the criteria of P ≤.05, 798 out of 15,280 features survived from this screening procedure and were recruited in the following stages of XGBoost algorithm.

The retrospective cohort was utilized for construction of the prediction model. This process was accomplished in two phases: (1) a data training algorithm to generate predictive estimates and (2) a calibration stage to produce risk scores for each patient.

For the first stage of data training, a supervised machine learning and data mining tool, XGBoost [35], was applied to develop the prediction model based on features survived in the univariate phase. Generally speaking, this algorithm generates an ensemble of classification trees and sums the scores in the corresponding leaves of each tree to calculate a final predictive estimate ŷ_i for the i- th instance (i=1,..., n), as demonstrated in equation 1, where each ƒ_k corresponds to an independent classification tree:

For this study, the depth of each tree was set to be 5 and K equaled 500. The model at the t- th iteration was trained to minimize the following objective, where l is a differentiable convex loss function that not only measures the difference between the target y_i and the prediction ŷ_i^(t-1) of the i-th instance at the t-1- th iteration but also takes the ƒ_t that improves the model most into account:

The term Ω penalizes the complexity of the regression tree functions to avoid issues of overfitting. The approximate greedy algorithm was adopted to split trees and sort and pick features on each node according to percentiles of distribution. Splitting points were selected to optimize purity at the next splitting level.

For the second phase of calibration, ŷ estimates were mapped to positive predictive values (PPVs) in the retrospective cohort. A PPV of a corresponding predictive estimate ŷ was defined as the proportion of incident hypertension events in the cohort having predictive estimates the same as or larger than this ŷ. Thus, PPVs could also be interpreted as risk scores that measured probability of receiving a new diagnosis of essential hypertension within the next 1 year among individuals having predictive estimates the same as or larger than this ŷ.

In the process of model evaluation, hypertension risk score expressed as PPVs were calculated and assigned to each individual in the prospective cohort. Following that, we further ranked individuals by their risk scores from low to high and allocated them into five distinct risk categories, manifesting individuals’ chance of developing essential hypertension within the next 1 year to be either very low, low, medium, high, or very high. Performance of the proposed 1-year hypertension risk prediction model was investigated carefully within each risk category in terms of sensitivity, specificity, and PPV. The receiver operating characteristic curve and the validated AUC value were derived to evaluate the performance of the derived prediction model.

For these highly weighted features, age-sex adjusted odds ratios (ORs) between cases and controls were calculated along with 95% CI using logistic regression. Distribution patterns of impactful risk predictors were explored and compared among different risk categories, uncovering predictors’ characteristics under our constructed prediction model. Moreover, survival analyses such as multivariable Cox regression were employed for the purpose of subpopulation comparison. Spearman rank correlations were performed to assess the correlation between socioeconomic features and the next 1-year risk of hypertension.

Age and gender were recognized as the most impactful demographic features in our hypertension risk prediction model. To follow the regulations of Health Insurance Portability and Accountability Act, all ages over 89 years were combined into one category to protect patients’ health information.

To maintain the consistency of this feature, other ages were aggregated into categories as well. In addition, this age categorization could help us capture natural differentials in age-related socioeconomic backgrounds such as work or retirement status and eligibility for pensions and medicare and investigate their impact on hypertension risk and incidence [36]. Therefore, from the very low risk to the very high risk category of hypertension, constituent ratios were calculated for four distinct age groups and compared with each other (Figure 4). Low-risk categories consisted of young people, whereas high-risk categories were concentrated with older individuals, confirming age group as a strong predictor of hypertension. Younger individuals (<35 years) comprised almost 89.27% (340,621/381,544) of the very low risk category. Middle-aged adults (35-49 years) were overwhelmingly dominant in the low risk and medium risk categories, with proportions of 51.04% (53,372/104,565) and 54.11% (53,792/99,415), respectively. On the contrary, individuals in the age range of 50 to 64 years occupied 34.57% (18,651/53,957) of the high risk population, and people aged ≥65 years constituted the largest subset (33.88%, 14,000/41,329) of the very high risk population (see Multimedia Appendix 4). Following that, we formed a subpopulation with people ≥65 years and found that our hypertension prediction model reached an acceptable to good discriminative ability with this subpopulation, with an AUC of 0.744 (see Multimedia Appendix 5). On the other hand, females were likely to have lower risk of hypertension than males, comprising majorities (58.25%, 341,072/585,524) of the three relatively low-risk categories (ie, very low, low, and medium) but decreasing to proportions of 46.89% (25,300/53,957) and 48.41% (20,008/41,329) in the high risk and very high risk groups (Figure 4). In summary, the very high risk population mainly comprised aged (>50 years) males, who occupied almost one-third of this category.

Most individuals in the very high risk population had a history of multiple chronic medical conditions. Most commonly CVDs, type 2 diabetes, and disorders of lipid metabolism affected 25.08% (10,367/41,329), 20.17% (8,336/41,329), and 8.76% (3619/41,329) of the population, respectively (see Multimedia Appendix 4), compared with only 0.65% (2473/381,544), 0.26% (1006/381,544), and 0.27% (1036/381,544) of the very low risk cohort.

To further explore real-time hypertension events in the very high risk category, we investigated time-to-hypertension curves in terms of disease subgroups (ie, prehypertension, type 2 diabetes, lipid disorders, mental illness, CKD, and CVDs) and compared them with curves generated from the very low risk cohort (Figure 5). In the very high risk category, 55.09% (5711/10,367) patients with CVD would receive diagnoses of hypertension in the next 1 year, and this probability stayed around 50% for patients with conditions of either type 2 diabetes, lipid disorders, mental illness, or CKD. Thus, a half of individuals with a chronic disease in the very high risk category would develop hypertension within the next 1 year, with CVD patients at the highest risk of hypertension during the next 1 year, whereas prehypertension patients had the lowest risk, implying different causal mechanism of hypertension under different disease scenarios at the very high risk stage. In contrast, in the very low risk category, more than 85% of patients were free from development of hypertension in the next 1 year, regardless of chronic conditions, where the survival curve dropped fastest for patients with prehypertension but went down slowly for the cardiovascular subgroup, opposite to the trends at the very high risk stage.

Apart from medications for CVD, type 2 diabetes, lipid disorders, and COPD, our prediction model also recognized 18 important medications prescribed for mental health diagnoses as powerful predictors of 1-year hypertension risk, which were mainly drugs prescribed for mood disorders such as depression, anxiety disorders, and schizophrenia disorders (see Multimedia Appendix 6). In our original HIE dataset from Maine, patient records related to mental health were initially blocked because of privacy protection issues. Therefore, our study utilized these consumed drugs as proxies to inspect the association between mental illness and risk of hypertension. People under treatment for depression, anxiety, and schizophrenia disorders were significantly enriched in the very high risk group for hypertension, with proportions of 22.21% (9181/41,329), 13.40% (5538/41,329), and 2.71% (1119/41,329), respectively, but much less prevalent in the very low risk group (see Multimedia Appendix 7).

Taking that most prevalent depression as an example, we further divided the prospective cohort into two subgroups of individuals who did and did not carry other chronic disease diagnoses and validated the association between mental disorders and hypertension risk accordingly.

By using multivariable Cox proportional hazards models, the contribution of depression to the hazard for hypertension was assessed after adjustment for age and gender (Figure 6). In the subgroup that comprised people with other chronic conditions, depression status revealed a relatively small effect on the 1-year hypertension hazard (HR: 1.1 [95% CI 1.1-1.2]), lower than both age and gender. However, in the subgroup of patients without other chronic physical conditions, we found that depression was associated with a two-fold hypertension risk (HR: 2.0 [95% CI 1.9-2.0]). Following that, time-to-hypertension curves were stratified by age (<65 and ≥65 years), depression status, and gender, in order of priority. Age, as the most impactful predictor, attained the highest HR (HR: 2.0 [95% CI 4.6-4.8]), whereas depression status was the second impactful feature. Therefore, as depicted in the two lowest survival curves in Figure 6, for individuals ≥65 years (mostly assembled in the very high risk category), a diagnosis of depression would double the 1-year hazard of hypertension even if there were no other chronic conditions, regardless of gender.

More severe health conditions are expected to consume more resources. Accordingly, clinical utilization indicators in our study revealed similar patterns to those highlighted diseases when compared across five risk categories. These utilization parameters such as outpatient visits, inpatient admissions, clinical cost, total number of consumed prescriptions, and total number of laboratory tests in the past year were lowest in the very low risk group but gradually increased from low to high risk groups, with the highest values in the very high risk population (see Multimedia Appendix 4).

On the basis of these findings, we focused compared patients in the very high risk and very low risk categories with respect to average clinical costs in the prior 12 months for 16 chronic disease subgroups, coordinated by their averaged number of chronic diseases (Figure 7). In the very low risk category, patients with prehypertension, CVDs, type 2 diabetes, lipid disorders, and mental disorders had relatively low prevalence (ie, small balls in Figure 7) and clustered in the lower left corner (ie, the low-cost and low-chronic-complexity area), very close to the largest bubble representing the reference subgroup formed by nonchronic disease individuals. However, in the very high hypertension-risk category, people usually suffered from multiple chronic conditions, resulting in more disease comorbidities and more clinical costs, bubbling toward the upper right corner. Moreover, in the very high risk category, the clinically costly CKD and chronic nephritis were recognized as the most severe chronic diseases on average, accompanied by six other chronic conditions, whereas CVD, type 2 diabetes, lipid disorders, and mental disease stayed in a moderate level of disease comorbidity and clinical cost.

Several interesting results were also found in socioeconomic features. In this study, most socioeconomic features were derived from ZIP code or county-based census and USDA data, and thus, they were recognized as community-level social and environmental indicators. The association between significant socioeconomic factors and the summarized hypertension risk scores were further investigated by Spearman rank correlations (Figure 8). Educational disparities revealed a unique and significant correlation with hypertension risk scores in our prediction model. The percentage of the low-education population (ie, a combination of the 18-24 year old population with less than high school graduate diploma and ≥25-year-old population with less than 12th grade diploma in the area) gradually increased from low- to high-risk categories and ultimately displayed a positive correlation (ρ=.05) with hypertension risk, occupying 8.73% (33,310/381,544) and 9.20% (3804/41,329) of the very low risk and the very high risk subsets (see Multimedia Appendix 8). In contrast, the averaged proportion of high-educated people that received college or associate’s degree or bachelor’s and higher degree grew in an opposite direction and showed a negative correlation (ρ=−.072) with hypertension risk, by attaining its highest percentage of 49.05% (187,148/381,544) in the very low risk group and decreased to its lowest value of 47.85% (19,776/41,329) in the very high risk population (see Multimedia Appendix 8).

More disparities were found in other social and economic indicators, both community-level persecutors of and defenders against hypertension (Figure 8). Those risk indicators with positive correlations (ρ>0) mainly comprised measures of nearby available food stores (eg, number of grocery and convenience stores per 1000 population and percentages of farmers’ markets selling fruit and vegetables and animal products within a certain region) and the percentage of people benefiting from Medicaid and Medicare type health insurance. On the other hand, the population-based defenders of hypertension (ρ<0) were identified as the median household income, percentage of people with low income and living far from supermarket or large grocery store (>1 mile for an urban area and >10 mile for a rural area), percentage of the population living within a half-mile of NAVTEQ-based parks, and percentage of the population covered by private health insurance, as the higher those values were, the lower the risk of hypertension would be in the population. In summary, in terms of social determinants, the very high risk population was likely to be occupied by low-income and low-educated people who benefited from public insurance and lived in an area near food stores but far from parks.

Most individuals in the very high risk group suffered from other medical conditions, and only 1.56% (645/41,329) had no other chronic medical conditions, whereas the majority (73.10%, 278,923/381,544) of the very low risk population had no other chronic disease diagnosis. Moreover, as the number of people with diagnostic chronic diseases increased, those with multiple chronic conditions (MCCs), also known as concurrent chronic conditions [38], became very common in the very high risk category (Figure 7), occupying almost a third (29.60%, 12,234/41,329) of the population, which also led to a dramatically increased burden of health care utilization and cost (Figure 5). In recent years, epidemiological studies indicated the high-level and continually increased prevalence of MCC worldwide. About a third of the US population suffer from MCC, the majority of which are working-age adults (45-64 years), whereas the prevalence is almost 80% among people ≥65 years [39,40]. Consistent with these findings, our very high hypertension risk category could be verified as an MCC population mainly consisting of elderly people with comorbidity of multiple chronic diseases. Hypertension is usually recognized as a major risk factor for cardiovascular and renal diseases. Conversely, within a complex reciprocal interrelationship, both essential and secondary hypertension can also be a consequence of chronic conditions [41,42], especially for elderly people [43]. Therefore, based on our prediction model, high risk of hypertension can be seen as the consequence of a long-term interplay of multiple chronic conditions via possible mechanisms such as polyunsaturated omega-3 free fatty acid deficiency, vascular endothelial dysfunction, immune dysregulation, and unhealthy lifestyle [44,45].

In addition, the definition of prehypertension varied in different studies from time to time [46-48]. In the sixth report of the Joint National Committee (JNC), prehypertension was defined as a systolic pressure of 130 to 139 mm Hg or a diastolic pressure of 85 to 89 mm Hg, whereas the updated seventh report of the JNC (JNC 7) defines prehypertension as a systolic pressure of 120 to 139 mm Hg or a diastolic pressure of 80 to 89 mm Hg. In our study, we derived the feature prehypertension from the ICD-9-CM category of 796.2 and the ICD-10-CM category of R03.0, defined according to JNC 7. Therefore, caution is needed when comparing our findings with those of previous studies, considering the different definitions of prehypertension.

The strength of association between mental disorders and subsequent onset of hypertension has been of interest for a long time. Much work has been done on specific psychological domains of depression, anxiety, impulsive eating disorders, and other mental disorders [49,50]. Furthermore, several studies have focused on the effects of mental disorders on subsequent hypertension [51,52]. Recognizing the impact of mental disorders on the onset of hypertension is clinically meaningful for several reasons. First, both the strength of effect and potential mechanisms of psychological and mental disorders to cause physical health conditions remain unclear [53,54]. Second, such association studies are driven by the ongoing worldwide increase of morbidity related to mental disorders.

In our study, the 1-year hypertension risk model and the depression-based survival analysis provided strong evidence of such an association, especially for individuals having no other chronic physical conditions. That is, about 32.17% (1226/3811) of elderly (≥65 years) patients having no chronic conditions but undergoing treatment for depression would develop hypertension within the next 1 year. EHR data in the state of Maine coding mental health diagnoses were masked for privacy protection; only consumed mental disorder drugs could be accessed as proxies. Therefore, as a limitation of the study, we cannot directly ascertain the effect of mental health disorders or the treatment thereof on the progression of hypertension. However, we can examine the relationship between mental illness diagnoses or therapeutics and hypertension development in several different ways according to the large amount of previous studies exploring this relationship [49-52,55-59]. First, mental illness itself could be a risk factor of hypertension, where possible mechanisms could relate to internal responses from altered sleep patterns, sympathoadrenal hyperreactivity, various neurotransmitter abnormalities, or altered inflammatory processes [55,56], as well as exposure to similar risk factors (eg, early childhood adversity) [57]. It is reported that depression can trigger deregulation in the sympathetic nervous system and hypothalamic-pituitary-adrenal axis, causing severe consequences such as metabolic syndrome and elevated risk of hypertension [59]. Second, mental illness medication could also independently increase the risk of hypertension through various drug-related side effects, which are connected to increasing rates of unhealthy lifestyles such as smoking, alcohol intake, physical inactivity, and drug addiction [45]. Certain drug treatments could also lead to the reprogramming of hypertension-related metabolic pathways. Previous cohort studies have found an increase in hypertension risk when tricyclic antidepressants are used; a possible mechanism is the effect of the aforementioned antidepressants on vagal control over the heart [58]. Future work is needed to address the causal relationship between mental illness and hypertension.

Social determinants of health are defined as “the structural determinants and conditions in which people are born, grow, live, work, and age” [60,61] and include social or environmental supportive factors of health, such as socioeconomic status, education level, employment and income, the physical environment, and access to health care. It has been reported that health disparities and inequalities in these social factors, especially education and poverty, accounted for over a third of total deaths in the United States [62]. Social determinants of health could be collected from global, national, and community levels, manifesting varying localized policies and actions on health promotion. Consistent with previous studies [63-66], our model demonstrated that the hazard of 1-year incident hypertension increased as the median household income and the percentage of high-educated population within a ZIP code area went down. In fact, lower education level could be translated as one of the initial driving and predisposing social features, leading to disparities of subsequent incident hypertension. That is, low education level could directly affect family income and ultimately increase the likelihood of having public health insurance, such as Medicaid or Medicare. In addition to those common social determinants of health, some lifestyle-related social factors were also identified in our study, namely, indicators of nearby accessible food stores and parks. The location of food stores and the types of products sold could affect underlying unhealthy diet habits, whereas environmental supportive facilities such as the concentration of parks or playgrounds in the living area may directly shape people’s physical activities and ultimately decrease hypertension risk. Another interesting finding was that, although low income level would increase hypertension risk overall, lower income populations had reduced hypertension risk if they lived far from food or grocery stores, implying an independent and critical impact of food consumption habits on hypertension regardless of income level. As the majority of the Maine population is white, our study failed to find any significant racial and ethnic disparities for hypertension risk, and it is possible that that our model is not as accurate with regard to individuals of other race or ethnicity. Another limitation of our study is that such community-level social determinants may potentially be less precise and could leave some useful social determinants uncaptured in the risk model.