Developing a Clinical Decision Tool

Left:  Gabrielle Johnson LPN Right: Dr. Andrew Schreiner MD (Primary Care and Internal Medicine)
Gabrielle Johnson LPN (left) and Dr. Andrew Schreiner M.D. (right)

 

A form of artificial intelligence known as deep learning could help primary care physicians treat chronic liver disease early.

MUSC Health primary care physician Andrew D. Schreiner, M.D., was recently awarded a five-year $865,000 grant from the National Institute of Diabetes and Digestive and Kidney Diseases to develop a clinical decision tool to help primary care physicians recognize and treat chronic liver disease in its early stages, when interventions are most effective.

Primary care physicians often see abnormal liver function tests for patients, but rarely intervene because little evidence currently connects these results to the development of specific chronic liver diseases.

“Due to the lack of current evidence, we don’t act upon them in a standardized or aggressive way in primary care,” said Schreiner. “And sometimes those abnormal liver tests are early clinical signals of underlying liver disease that may be problematic in the future.”

Chronic liver disease is now the twelfth leading cause of death in the U.S. according to the Centers for Disease Control and Prevention, and the rate of chronic liver disease is on the rise. For example, non-alcoholic fatty liver disease, now affects 30 percent of people in developed countries, making it the most common cause of chronic liver disease in the Western world.

Too often, patients learn they have chronic liver disease only after it has become very advanced and treatment options are limited. Earlier identification of disease could lead to earlier treatment and better outcomes.

For instance, patients with hepatitis C could be treated with newly available, effective medication that can rid them of the virus. Patients with non-alcoholic fatty liver disease could be counseled
on diet and weight loss and advised to begin statins.

Schreiner will review the records of 11,000 primary care patients with abnormal liver function tests and track which of those patients went on to develop a chronic liver disease. Although “normal” liver function tests may be similar, abnormal ones vary widely, and that variation could be useful in predicting which chronic liver disease a person is likely to develop, if any.

An experienced team of bioinformaticists will look for correlations between given “phenotypes” of results and specific chronic liver diseases and will use those, along with clinical variables and demographic information available in the electronic health record, to develop predictive analytics.

The team of bioinformaticists who have successfully created predictive models for kidney transplant will use deep learning to find the correlations between test result phenotypes and specific chronic liver diseases. They will then design a decision support tool that would reside on the electronic health record and provide treatment recommendations to physicians that are tailored to each patient’s specific liver function test results.

If the team is successful in developing a decision support tool during the grant funding period, the next step will be to trial it in a large primary care practice.

Predicting epilepsy surgery outcomes with deep learning

Using deep learning, MUSC Health neurologists have developed a new method that may one day help both patients with medication-refractory epilepsy and their physicians weigh the pros and cons of brain surgery.

Although brain surgery is often recommended to patients who do not respond to medication, many hesitate, in part due to the operative risks and in part due to limited success.

To overcome this, Leonardo Bonilha, M.D., Ph.D., and his team searched for a better way to predict which patients are likely to be seizure free after surgery. The team turned to deep learning due to the massive amount of data analysis required.

“In this study, we incorporated advanced neuroimaging and computational techniques to anticipate surgical outcomes, with the goal of enhancing quality of life,” explains Neurology Department Chief Resident Ezequiel Gleichgerrcht, M.D.

Post-Surgery Outcomes Accurately Predicted

A pair of pie charts showing the difference between post-surgery outcome predictions using clinical variables (evenly split) versus using deep learing (79 to 88 percent accuracy).

The whole-brain connectome, the key component of this study, is a map of all physical connections in a person’s brain. The map is created by in-depth analysis of diffusion magnetic resonance imaging (dMRI), which patients receive as standard-of-care prior to surgery. The neurologists used deep learning to examine the connectome, allowing for patterns to be automatically learned.