Dam Data: Health Systems, Machines, And Learning

April 9, 2018

Contributed Commentary by Michael Pencina

April 9, 2018 | Too often evidence in health care resembles raindrops in a storm. Whether or how much you get wet depends on where you stand metaphorically speaking. How far and how deeply a clinical insight penetrates can have little to do with the strength of the evidence at its foundation.

Increased use of electronic health records (EHRs) and other digital data has brought what amounts to a deluge of potential evidence, one that never ends. A “learning health system” can create the equivalent of a hydroelectric dam – a place to catch all of those raindrops in one place and release them in a coordinated, powerful way. As the National Institutes of Health describes it, a learning health system “creates a continuous cycle or feedback loop in which scientific evidence informs clinical practice while data gathered from clinical practice and administrative sources inform scientific investigation.”

Given the amounts of data available, machine learning is quickly becoming indispensable to making that loop from science to clinical practice. Machine learning allows practitioners to reach beyond their own experiences and access the sum of experiences of all patients of all physicians in the health system.  It mimics how physicians think about treatment options, but at a scale that can only be achieved with computing.

But machine learning is not sufficient on its own.  Clinical expertise, study design, and deep understanding of the data to be analyzed are as critical as the machine learning methods to be applied. This makes a university health system an ideal proving ground. Proximity to the university’s statistical, computer science, engineering, and clinical researchers encourages the fusion of expertise that must occur.

Among the machine learning projects we are undertaking, perhaps the closest to mimicking the traditional use of evidence is a decision support tool to help clinicians managing diabetes patients’ glucose levels. There are a number of established and novel treatment options for patients with high hemoglobin A1c levels; patients respond differently to each. It’s not uncommon for a patient to see no decrease in Hb A1c levels even after trying two or three treatments. With machine learning, we are able to scale a physician’s decision-making process. Rather than rely on one physician’s anecdotal knowledge and experience with similar patients, we can harvest the data of all patients with high Hb A1c with clinical and treatment characteristics similar to our patient at hand.

We use a deep recurrent neural network model in which the algorithm includes patients with two elevated levels prior to a treatment change and one re-check after a treatment change. That, however, excludes patients with no change in treatment—an important population. Drawing on faculty expertise in causal inference, we use patient characteristics to create longitudinal matches between patients with and without medication change. Our machine learning model is applied to the sample created to estimate the expected Hb A1c treatment response according to the medication that would be prescribed, or no medication change.

In another project involving diabetes patients, we are exploring a deep learning model to predict their risk of a wide range of complications. This Deep Poisson factor model takes advantage of characteristics available through advanced machine learning—non-linear analysis, search of a complex model space, imputation through a correlation structure, a single model for all outcomes—that differ from the more traditional LASSO regression we are using as a comparison.    

The power of predictive modeling through machine learning has many applications. Machine learning can improve care beyond what is captured in EHRs as well. Anyone who has purchased a flat screen TV knows that their display involves thousands of dots that come together to form an image. A deep learning algorithm adept at learning by reading these pixels can process visual information including medical imaging. We are training this kind of algorithm to analyze various types of medical images (for example, invasive coronary angiograms) to search for the ability to determine the significance and location of various forms of disease (for example, obstructive coronary stenosis).  A model built and trained on a consensus opinion of practitioners in a health system—or even across several health systems—would likely have better, and certainly more consistent, performance given what we know about inter-rater reliability. Eventually, a system utilizing such an algorithm could interact with physicians as they perform the procedure, providing real time feedback to ensure focus on the most significant areas of an image.

What’s certain is the deluge of healthcare data will continue to intensify. What these examples demonstrate is the challenge presented to clinicians, statisticians, bioinformaticists, engineers, and computer scientists is to work together to strengthen tools built on artificial intelligence to take advantage of this flow of evidence. When they do, the result will be health care that’s more effective and more precise for each patient.

Michael Pencina, PhD, is faculty associate director and director of biostatistics for the Duke Clinical Research institute and a professor of biostatistics and bioinformatics at the Duke School of Medicine.