Natural Language Processing

Opportunity 

Based on nationwide guidelines, there are specific medications that should be used to treat patients suffering from heart failure with a reduced ejection fraction (EF), or a reduced pumping function, causing symptoms of fatigue, shortness of breath and swelling. However, EF values, measured in percentages, are typically buried in doctors’ notes and test results, making it difficult for other physicians to easily obtain the information they need to treat their patients with the correct medications.

researchers on computers

Solution 

OSF HealthCare developed the Heart Failure Council to focus on reducing mortality and re-hospitalization rates among this population using innovative methods and collaborative thinking. Working with the  Advanced Analytics team, a part of OSF Innovation, a Natural Language Processing (NLP) model was developed to intuitively read and pull all EF numbers for heart failure patients into the electronic medical record (EMR) system, giving clinicians an easily accessible historical view of a patient’s heart health to make an informed decision on medication prescriptions.

Impact

Before the NLP project, there were 56,000 ejection fractions as whole numbers in the EMR for OSF HealthCare. These were from echocardiograms (echoes) previously performed at OSF, dating back to 2014. After the NLP program was initiated, an additional 152,000 EF values were mined from echoes taken all the way back to 2007, giving clinicians the ability to assign a reduced, borderline or preserved phenotype to twice as many patients in the heart failure registry. It also identified 2500 additional people with low EF values that were not included in the registry since they had no diagnosis of heart failure. As a result, clinicians have been able to easily pinpoint heart failure patients with historically low EF values and ensure those individuals receive the medications they need. 

“A doctor can now pull up a patient’s medical chart and see within the Results Review tab that their EF was measured seven times over their life. The provider then has a longitudinal view of the patient’s heart health and can make an informed decision on medication prescriptions.”

- Dr. Parker McRae, change agent, Cardiovascular Quality Improvement, OSF HealthCare

Download PDF

Easily Tracking EF Values

Heart failure remains a growing epidemic in the United States with an estimated one in five Americans developing some form of heart disease in their lifetime. While it’s usually incurable, individuals with the condition can live full healthy lives with the right medications and healthy lifestyle changes.

In an effort to keep heart failure patients healthy, OSF HealthCare developed the Heart Failure Council to focus on reducing mortality and hospital readmission rates among this population using innovative methods and collaborative thinking. With that came the need to develop a better way to identify individuals suffering from heart failure with a reduced ejection fraction (EF), or a reduced pumping function.

A low EF indicates a patient’s heart isn’t pumping out enough blood to the rest of the body, causing fatigue, shortness of breath or swelling. Nationwide guidelines recommend that heart failure patients with a history of any EF at or below 40% should be prescribed certain medications to help their hearts function better. However, finding an individual’s EF metric is difficult as it’s typically buried in doctors’ notes and test results.

The OSF HealthCare Heart Failure Council worked with the Advanced Analytics team,  a part of OSF Innovation,  to develop a Natural Language Processing (NLP) model that intuitively reads and pulls all EF numbers for heart failure patients into the electronic medical record system which clinicians use daily.

Building an NLP Program

Physicians historically document the EF of heart failure patients in written notes and tests from echocardiograms (echoes) that are saved in a large database. The problem is that the information is mixed in with a lot of other internal and external data, so there was no easy way to find out whether a patient had been diagnosed with reduced EF in the past.

“Unfortunately, when a patient’s EF is found to be in the normal range, physicians don’t continue treating them with certain heart medications,” said Dr. Parker McRae, a change agent for Cardiovascular Quality Improvement at OSF HealthCare. “Importantly however, that person’s EF may be normal only because of the drug therapy they were receiving. Having an easily displayed historical view of a patient’s EF over time will help clinicians make better decisions on treatment.”

Working with Dr. McRae, the Advanced Analytics team developed an NLP model that was trained on a sample of more than 1,000 hand-annotated echo notes. It also uses a machine learning technique known as “named entity recognition” designed as a convolutional neural network to find EF scores. Those values are then extracted, according to rules defined by Dr. McRae and are stored in the clinical databases of OSF HealthCare.

In order to validate model performance, Dr. McRae dove into a sample of 500 echocardiograms and manually separated the EF scores from the text. The trained NLP model was then applied to the same 500 notes that Dr. McRae abstracted. The model exactly matched Dr. McRae about 97% of the time and the identification of an EF at or less than 40% matched 99.7% of the time.

“The NLP model is essentially an automated and highly accurate replica of Dr. McRae’s decision-making capabilities for the specific task of abstracting EF scores from echocardiogram notes,” said Jason Weinberg, a data scientist with the Advanced Analytics team.

This validation process has been completed twice, ensuring model accuracy remains clinically relevant. The NLP model has now been applied to all OSF echoes, unlocking data that is critical to a heart failure patient’s health.

Value and Impact

Before the NLP project, there were 56,000 ejection fractions as whole numbers in the EMR for OSF HealthCare . These were from echoes previously performed at OSF, dating back to 2014.

After the NLP program was initiated, an additional 152,000 EF values were mined from echoes taken all the way back to 2007, giving clinicians the ability to assign a reduced, borderline or preserved phenotype to twice as many patients in the heart failure registry.

It also identified 2500 additional people with low EF values that were not included in the registry since they had no diagnosis of heart failure. There are now 25,000 people in the heart failure registry.

As a result, 11 OSF facilities charged with identifying heart failure patients with low EF scores have seen the use of guided direction medication treatment (GDMT) increase by nearly 100% for two primary heart failure medications. There’s been a nearly 600% upturn in clinicians prescribing all three GDMT medication classes in combination.  This is expected to improve the outcomes of these patients.

What’s next?

Advanced Analytics isn’t finished training the NLP program to extract EF numbers. The next phase will pull in physician notes that are on MUGA scans, cardiac MRIs and cardiac CTs to get an even more  EF values.

The team is also working to predict the likelihood of a heart failure patient to be admitted to the hospital. This will be based on normal office visits with a physician. If this information can be determined  ahead of time, OSF HealthCare can provide the resources necessary to help stop even the first hospital admission from happening.