Machine learning can bring more intelligence to radiology
Machine learning is emerging as one of the key hopes to change the practice of radiology—the opportunity seems ripe, with rising calls for radiologists to demonstrate increased quality and more value, even as technology yields bigger datasets and more complexity.
But exactly how machine learning will impact the radiology profession—and healthcare in general—remains to be seen. It will just take time and experimentation with machine learning, some say.
Keith Dreyer, DO, likens the machine learning revolution to the promulgation of electricity, which originally was used simply for lighting, but eventually ushered in a host of helpful inventions—washing machines, dishwashers, air conditioners, televisions, computers—that were previously unimaginable.
“Once you start to make machines think, taking data and performing predictive analytics, things will happen that are beyond human capability and current imagination,” said Dreyer, vice chairman of radiology at Massachusetts General Hospital and associate professor of Radiology at Harvard Medical School. “So if you could predict a group of patients that are likely to have a positive CT of the brain before it was performed, think of the advantage that would be.”
How it can help
As computers outperform humans at complex cognitive tasks, machine learning has enormous potential to enhance diagnostic accuracy, predict prognosis and, ultimately, improve patient outcomes.
J. Raymond Geis, MD, Department of Radiology at University of Colorado described how researchers at Mayo Clinic use a machine learning algorithm on a very well-defined problem of a brain tumor called glioblastomas. Different types of glioblastomas have different genetic abnormalities, and based on those genetic abnormalities, physicians treat them differently. However, radiologists looking at images of glioblastomas can’t predict which genetic variation they are. But Mayo’s machine learning program can look at this very specific clinical problem and identify genetic abnormalities.
“The advantage for radiology is in these small, well-defined clinical situations where the machines can get more information from the images than two human eyes can distinguish,” Geis says.
While still in its early stages, a critical task of machine learning in radiology is to extract more knowledge from data. “In medical imaging, we’ve dramatically increased the capability of visualization of the data, but what we haven’t improved on in the last decades is to create quantifiable data coming out of those modalities,” Dreyer says. “Once we have algorithms that are capable of doing that in an automated sense with high reliability, the output from diagnostics is going to be more consistent—the result will be much stronger predictive capabilities for diagnostics in precision care.”
Many products, little validation
But various challenges first must be overcome to achieve widespread use of machine learning. For starters, while there are thousands of different machine learning programs from hundreds of vendors on the market, each one is designed for a very specific clinical situation. But in daily practice, physicians are seeing hundreds of thousands of kinds of pathology, and it’s simply not feasible to invest in thousands of narrowly focused machine learning software programs.
That diversity of offerings is challenging for hospitals to manage. Today, many health systems are consolidating IT purchases, a trend that runs counter to that posed by machine learning offerings.
Some experts caution against being an early adopter of machine learning, contending that the technology has yet to be validated. “There is way too much emphasis on a particular capability,” said Paul Chang, MD, professor and vice chairman of radiology informatics at the University of Chicago School of Medicine. “I don’t like approaches or disruptions that concentrate on technical capability. I’d much rather it be driven by use case.”
Current use cases are immature or not compelling because of the fundamental challenge of deep learning. The vast majority of deep learning algorithms require supervised training with a data set. For example, if you want to create a deep learning program that can detect lung nodules and determine whether they’re cancerous, you have to train it by sending it a batch of images that you identify as having cancer and a batch of images that you identify as not having cancer.
Those vetted data sets have to come from production systems. “We don’t have research data sets,” Chang says. “The huge barrier to deep learning is that we don’t have training data sets.”
Building capability beyond EMR
Right now, hospital IT organizations are EHR-centric, which presents another obstacle to the implementation of machine learning. “EHRs are not currently equipped to handle broad scale AI integration,” said Dreyer. “We do, however, have a vehicle for integration of predictive analytics in our current evidence-based Clinical Decision Support interface. We plan to use this vehicle for our AI integration as well.”
Experts agree that a good deal of fundamental framework is required before machine learning can truly take hold in healthcare.
Chang believes hospital CIOs would be wise to take a “hedge strategy.” He sees a huge economic risk to diving into deep learning too early. Moreover, he adds, deep learning is just one of many cybernetic decision support capabilities—Bayesian networks, analytics, big data, registries, to name a few others—that CIOs need to be prepared to manage.
However, EMR systems were designed to help clinicians from an operational standpoint and were not designed to address this new need. “EMR is not an architecture,” Chang says. “It’s one of many components of a true enterprise service bus that allows us to support modern needs.”
The challenge for CIOs is to figure out now how to get one’s EMR-centric IT infrastructure to be able to feed and consume these kinds of cybernetic services at scale in a HIPAA-compliant way. “If I were a CIO, I’d be thinking: I better start building the capabilities in my existing IT stack so that when deep learning is validated and becomes real, I’m ready to use it and exploit it,” Chang says.
For hospital IT execs, that means determining a way—at scale and compliant with HIPAA rules—to extract all the data that currently stands in various silo datasets and correlate it with outcomes to feed these new systems.
The use of machine learning in healthcare is already happening, Dreyer contends. In radiology, machine learning algorithms are already detecting pulmonary nodules, diagnosing polyps and screening for breast cancer. However, many more algorithms are on the horizon.
Once economic and IT barriers are addressed, machine learning has the potential to dramatically improve the ability of physicians to establish a prognosis. For example, it could be used to look at all lung cancer patients and then correlate them with their lab values, genetic profiles and diagnostic images to find patterns that help doctors.
But according to Geis, we might also see applications of machine learning where the data that’s generated helps to measure value in medicine. Machine learning has the potential to look at medical data—EHR data, financial data, measurement of outcomes—and search for patterns based on individual providers or groups of providers. That would enable algorithms to tease out individual contributions to care.
For example, an algorithm might assess the effect of two doctors by looking at patterns that happen to patients after they are treated. It has the potential to identify factors like, “everyone who goes to Doctor A seems to do a little better or their care costs a little less,” says Geis. “Once people start looking at it that way, this will be one of the biggest explosions of machine learning.”