Electronic health records can help identify patients with hypertension, according to a new study published in the Journal of the American Medical Informatics Association (JAMIA).
Hypertension affects one-third of Americans and contributes to one in six adult deaths in the United States, so it’s important to identify those who have it. The researchers, from Vanderbilt University and elsewhere, developed and evaluated the performance of different algorithms and EHR data types to identify individuals with and without hypertension.
They reviewed the EHRs of 631 individuals followed at Vanderbilt for hypertension status, using four phenotyping algorithms of increasing complexity. The input categories included ICD-9 codes, medications, vital signs, narrative text search results and medical language system concepts extracted through natural language processing.
The researchers discovered that even the simple algorithms performed well, and that those that were more complex, with combinations of feature categories, performed even better.
Blood pressure measurements, despite being the basis for determining hypertension clinically, performed worst of all categories for the identification of hypertensive individuals from EHR data. The researchers surmised that this was likely due to issues such as treatment reducing blood pressure within normal range, treatment often starting outside the EHR dataset and the many non-hypertensive causes of high blood pressure within the EHR. They also found that using readings from both inpatient and outpatient settings outperformed readings from outpatient settings only. Most of the algorithms could be replicated in another facility.
“[W]e can identify hypertensive individuals with high recall and precision by combining EHR data sources," the researchers said. "Even simple combinations of elements from different categories are statistically significantly better than current simple ICD-9 code count thresholds. ... Models and features based on structured EHR fields are more portable than text-based features, especially regular expression-based features. The best phenotyping algorithms have broad potential applicability.”
To learn more:
- here's the study abstract