The vision: Data-informed care for better outcomes and more empowered clinicians
Hypertension, or High Blood Pressure (HBP), affects nearly half (47%)1 of the U.S. population and contributes to almost 500,000 deaths per year.2 For this pioneering health system, it’s a sobering statistic that inspired a new initiative: tap into the power of Machine Learning (ML) and help patients get their HBP under control.
Hypertension is difficult to control for several reasons: Providers must follow population-specific, not patient-specific, guidelines when prescribing medications. Designing a within-guideline regimen is complex, and there are many options for treatment with more than 70 medications available — each with different recommended dosages and combinations from different classes. This makes the number of possible prescriptions for each patient astronomical. For our client, this complexity had resulted in 60% of its patient population not having their hypertension controlled.
To provide more data-informed and targeted treatment, our teams got to work on an ML engine that would provide recommendations based on patients' unique symptoms, comorbidities, current medication(s) and other possible medications.
A healthy foundation
First, Insight engaged with the health system’s clinicians and surgeons to understand workflow. We also conducted a viability assessment of the current state of hypertension treatment and potential use of ML in the environment. Insight and the clinic's technical team identified technology, general acumen and scalability in the data science as key hurdles.
Our teams then designed and assembled a sizeable Electronic Health Record (EHR) data model compatible with the requirements of ML algorithms. EHR data, including vitals, medications, comorbidities, laboratory results, ejection fraction, encounters and census data, were incorporated into the data model. Vitals, medications and encounters were also processed to incorporate temporal changes of each into the data model. For example, if a patient had prior blood pressure measurements, blood pressure changes and trends were calculated. Similar calculations were done for the most common patient-hospital encounters.
To improve predictive accuracy and inform clinical support, nearly 800 patient data points or patient features were used or engineered for every patient encounter. Our teams ensured that no Personally Identifiable Information (PII) was downloaded to local computers.