Use of machine learning to predict hypertension-related complication outcomes of varying severity

Abstract

Objective A challenge in hypertension-related risk management is identifying which people are likely to develop future complications. To address this, we present administrative-claims based predictive models for hypertension-related complications. Materials and Methods We used a national database to select 1,767,559 people with hypertension and extracted 112 features from past claims data based on their ability to predict hypertension complications in the next year. Complications affecting kidney, brain, and heart were grouped by clinical severity into three stages. Extreme gradient boosting binary classifiers for each stage were trained and tuned on 75% of the data, and performance on predicting outcomes for the remaining data and an independent dataset was evaluated. Results In the cohort under study, 6%, 17%, and 7% of people experienced a hypertension-related complication of stage 1, stage 2, or stage 3 severity, respectively. On an independent dataset, models for all three stages performed competitively with other published algorithms by the most commonly reported metric, area under the receiver operating characteristic curve, which ranged from 0.82-0.89. Features that were important across all models for predictions included total medical cost, cost related to hypertension, age, and number of outpatient visits. Discussion The model for stage 1 complications, such as left ventricular hypertrophy and retinopathy, is in contrast to other offerings in the field, which focus on more serious issues such as heart failure and stroke, and affords unique opportunities to intervene during earlier stages. Conclusion Predictive analytics for hypertension outcomes can be leveraged to help mitigate the immense healthcare burden of uncontrolled hypertension.

Real world hypertension treatment patterns analysis identifies putative gaps and risk-associated patterns

“…For such interventions, high risk ATL_THREE_ROTO patients may be initially targeted based on their CCI score/existing co-morbidities or persistence of ATL_THREE_ROTO status over multiple years or both, and further prioritized - to potentially increase the likelihood of successful interventions - by focusing initially on those patients having relatively low CDS scores (i.e., less complicated treatment regimens). Outside of targeting patients by ROTO status, it may also be used to monitor efficacy of HTN care programs overall – wherein high risk patients are identified via predictive modeling en masse [16] and patients monitored for their ROTO status and/or reduction of ROTO count (if initially belonging to ATL_THREE_ROTO class). Further work is needed to validate these notions in pilot programs that monitored these metrics alongside other standard measures (medication adherence, stability of HTN readings & self-monitoring, quality of life, weight/BMI management measures) for risk reduction in HTN-specific care programs.…”

“…HTN complications were grouped as stage 1/stage 2/stage 3 complications based on underlying medical claim codes. The underlying medical codes used for grouping complications based on severity and overall modeling strategy details are described in an accompanying manuscript [16]. The most relevant details for this report are outlined below.…”

Population level risk stratification of hypertensive patients by predictive modeling

“…To help with early identification of "high-risk" HTN patients from administrative claims, researchers have developed models for predicting HTN patients at risk of future (HTN-related) complications [6 7], including those that are grouped by severity [8]. Multiple groups have also performed treatment pathways/outlier analyses at scale to characterize prescription patterns among new HTN patients [9 10], compare various first line treatments for efficacy/safety [11], and identify future riskassociated treatment patterns among new vs all HTN patients [12].…”

“…For identification of HTN-related medications, following anatomical therapeutic classification (ATC) system codes were used: C02, C03 except C03C, C04AB, C07, C08, C09 based on an independent report [16]. HTN-diagnosis codes were as previously described [8]. SQL and Python scripts were used for identification of study cohort, feature generation, predictive modeling analyses, and characterization of model outputs/predictions described in further detail below.…”

“…After stratifying above study cohort by atl_three_roto status in 2015, logistic regression models were generated to predict among atl_three_roto vs not_atl_three_roto patients in 2015 who will develop "any stage" HTN complications (outcome variable) in 2016. Patients with "any stage" HTN complications were identified from 2016 administrative claims with ICD-9/10 diagnostic codes corresponding to increasing levels of severity (stage 1, stage 2, or stage 3 complications) across different target organ systems, as described in a previous report [8]. Input features (see Supplementary table S2) for the two predictive models were calculated using 2015 administrative claims data and included: demographic features, disease diagnosis groups, medication types, and 3 categorical features (representing distinct Dx claim dates with HTN and HTN Rx counts).…”