Identification of Hypertension in Electronic Health Records Through Computable Phenotype Development and Validation for Use in Public Health Surveillance: Retrospective Study.

BACKGROUND: Electronic health record (EHR) systems are widely used in the United States to document care delivery and outcomes. Health information exchange (HIE) networks, which integrate EHR data from the various health care providers treating patients, are increasingly used to analyze population-level data. Existing methods for population health surveillance of essential hypertension by public health authorities may be complemented using EHR data from HIE networks to characterize disease burden at the community level. OBJECTIVE: We aimed to derive and validate computable phenotypes (CPs) to estimate hypertension prevalence for population-based surveillance using an HIE network. METHODS: Using existing data available from an HIE network, we developed 6 candidate CPs for essential (primary) hypertension in an adult population from a medium-sized Midwestern metropolitan area in the United States. A total of 2 independent clinician reviewers validated the phenotypes through a manual chart review of 150 randomly selected patient records. We assessed the precision of CPs by calculating sensitivity, specificity, positive predictive value (PPV), F1-score, and validity of chart reviews using prevalence-adjusted bias-adjusted κ. We further used the most balanced CP to estimate the prevalence of hypertension in the population. RESULTS: Among a cohort of 548,232 adults, 6 CPs produced PPVs ranging from 71% (95% CI 64.3%-76.9%) to 95.7% (95% CI 84.9%-98.9%). The F1-score ranged from 0.40 to 0.91. The prevalence-adjusted bias-adjusted κ revealed a high percentage agreement of 0.88 for hypertension. Similarly, interrater agreement for individual phenotype determination demonstrated substantial agreement (range 0.70-0.88) for all 6 phenotypes examined. A phenotype based solely on diagnostic codes possessed reasonable performance (F1-score=0.63; PPV=95.1%) but was imbalanced with low sensitivity (47.6%). The most balanced phenotype (F1-score=0.91; PPV=83.5%) included diagnosis, blood pressure measurements, and medications and identified 210,764 (38.4%) individuals with hypertension during the study period (2014-2015). CONCLUSIONS: We identified several high-performing phenotypes to identify essential hypertension prevalence for local public health surveillance using EHR data. Given the increasing availability of EHR systems in the United States and other nations, leveraging EHR data has the potential to enhance surveillance of chronic disease in health systems and communities. Yet given variability in performance, public health authorities will need to decide whether to seek optimal balance or declare a preference for algorithms that lean toward sensitivity or specificity to estimate population prevalence of disease.

Measuring agreement between decision support reminders: the cloud vs. the local expert.

BACKGROUND: A cloud-based clinical decision support system (CDSS) was implemented to remotely provide evidence-based guideline reminders in support of preventative health. Following implementation, we measured the agreement between preventive care reminders generated by an existing, local CDSS and the new, cloud-based CDSS operating on the same patient visit data. METHODS: Electronic health record data for the same set of patients seen in primary care were sent to both the cloud-based web service and local CDSS. The clinical reminders returned by both services were captured for analysis. Cohen's Kappa coefficient was calculated to compare the two sets of reminders. Kappa statistics were further adjusted for prevalence and bias due to the potential effects of bias in the CDS logic and prevalence in the relative small sample of patients. RESULTS: The cloud-based CDSS generated 965 clinical reminders for 405 patient visits over 3 months. The local CDSS returned 889 reminders for the same patient visit data. When adjusted for prevalence and bias, observed agreement varied by reminder from 0.33 (95% CI 0.24 - 0.42) to 0.99 (95% CI 0.97 - 1.00) and demonstrated almost perfect agreement for 7 of the 11 reminders. CONCLUSIONS: Preventive care reminders delivered by two disparate CDS systems show substantial agreement. Subtle differences in rule logic and terminology mapping appear to account for much of the discordance. Cloud-based CDSS therefore show promise, opening the door for future development and implementation in support of health care providers with limited resources for knowledge management of complex logic and rules.