Improving hypertension diagnosis through machine learning in a national general practice database

Background
Millions of people with hypertension remain undiagnosed and untreated, resulting in avoidable consequences. We aimed to use machine learning to improve the identification of patients at risk of hypertension by using electronic medical records (EMR).

Methods
Cross-sectional study using data from MedicineInsight, a database containing de-identified EMR of 1.9 million patients attending 650 general practices across Australia. Manual algorithms were used to identify patients with a newly recorded hypertension diagnosis between 2016-2018 in 735,000 regular adults without a history of hypertension: 1) two or more records of elevated blood pressures (BP ≥140/90); or 2) diagnosis, encounter reason, or prescription reason of “hypertension”; or 3) prescription for anti-hypertensive therapy preceded by an elevated BP. Using a dataset where the above information was removed, deep learning techniques (a subset of machine learning) were employed to determine whether patients with hypertension could be identified based on other information in the EMR.

Results
The manual algorithm identified 119,230 adults with a newly recorded hypertension diagnosis (average age 56.4 ±15.8 years; 53.0% female). The sensitivity and specificity of the deep learning techniques were 81.0% and 71.3%, respectively. Preliminary analysis was able to identify complex interactions in the EMR to identify patients with hypertension.

Conclusions
Using routinely collected data from primary healthcare settings, deep learning techniques are able to identify patients who may have hypertension using non-traditional variables.

Key message
Machine learning could support the identification of patients at risk of hypertension by using EMR and flag these patients to general practitioners for further investigation.