Last February, IBM’s Watson supercomputer captured the world’s imagination when it bested a pair of Jeopardy champions. However, the game show represents the tip of the iceberg as far as meaningful applications of the technology. Eliot L. Siegel, MD, professor of radiology and vice chair of imaging informatics at the University of Maryland School of Medicine in Baltimore, has been involved with Watson’s development for two years and offers a sneak peek at how the supercomputer may change the face of medicine.
Q: Can you describe how you first became involved with IBM and Watson?
Siegel: As part of my research, I worked with an IBM researcher looking at image segmentation and image recognition and asked about the Deep Blue chess program. The reinvented team was working on a project to challenge the best Jeopardy players of all time. It entailed developing software that could respond generically to any question and come up with an answer associated with a knowledge base. I realized the approach had tremendous potential in medicine.
Q: How has the research proceeded?
Siegel: IBM researchers started working with me and my Maryland Imaging Research and Technology Laboratory (MIRTL) on medical applications. IBM began augmenting Watson’s database with medical textbooks and journal articles, and as its understanding became deeper, we tested the system with questions from internal medicine board exams. The next step was testing its ability to take signs and symptoms and formulate a differential diagnosis, refine that diagnosis and start suggesting treatment options.
Q: Were there challenges in Watson’s medical education?
Siegel: One challenge is taking a good question-and-answer engine that uses different algorithms and translating that into something that works from a practical perspective in a healthcare setting. Although Watson has tremendous encyclopedic knowledge of medicine, it needs practical experience.
We can feed Watson practical experience by providing it access to large numbers of EMRs from different databases and sources. There’s potential for Watson to use natural language processing (NLP) to acquire information from millions of patient records in an anonymous fashion.
Q: What type of Watson-powered applications might we see in radiology?
Siegel: When I’m looking at a new imaging study, I typically need more information than I’m provided. It isn’t always practical to spend 10 to 20 minutes reading through the medical record. At an academic institution, a resident or fellow can review the chart.
However, non-academic radiologists don’t have that luxury. They could utilize Watson to look through a patient record and summarize what’s important.
Q: What supporting infrastructure is necessary to realize Watson’s potential?
Siegel: Watson faces many challenges. Most places capture patient progress notes and discharge summaries in an unstructured, non-digital format. Watson learns best from structured information such as lab values and hemodynamic data.
Another challenge is the often contradictory nature of unstructured information. Anyone can start a problem list, but there is no sheriff to remove outdated or false data. In medicine, clues may be incorrect, information may be contradictory, there may be multiple answers to a given question or multiple possible diagnoses to a given set of medical information. Although there are multiple databases that Watson can access, there aren’t standards for interacting with those databases. However, if Watson is even partly successful, it has the potential to substantially change the face of medicine.
Q: What imaging applications might we see in the future?
Siegel: There may be applications in image analysis and image segmentation, such as a melding of technologies combining CAD with Watson’s abilities to summarize, synthesize and suggest diagnostic options.