A man walks into a doctor’s office for a CT scan of his gallbladder. The gallbladder is fine but the doctor notices a saclike pocket of fluid on the man’s pancreas. It’s a cyst that may lead to cancer, the doctor tells him, so I’ll need to cut it out to be safe.
It’ll take three months to recover from the surgery, the doctor adds—plus, there’s a 50 percent chance of surgical complications, and a 5 percent chance the man will die on the table.
An estimated 800,000 patients in the United States are incidentally diagnosed with pancreatic cysts each year, and doctors have no good way of telling which cysts harbor a deadly form of cancer and which are benign. This ambiguity results in thousands of unnecessary surgeries: One study found that up to 78 percent of cysts for which a patient was referred to surgery ended up being not cancerous.
Now there’s a machine learning algorithm that could help. Described today in the journal Science Translational Medicine, surgeons and computer scientists at Johns Hopkins University have built a test called CompCyst (for comprehensive cyst analysis) that is significantly better than today’s standard-of-care—a.k.a. doctor observations and medical imaging—at predicting whether patients should be sent home, monitored, or undergo surgery.
“We are extremely excited about the results of this,” said senior author Anne Marie Lennon, director of the pancreatic cyst program at the Johns Hopkins Kimmel Cancer Center, at a press conference about the study. She expects to offer the test to Hopkins patients within 6 to 12 months and hopes to make it commercially available following a larger, prospective clinical trial.
The vast majority of pancreatic cysts are benign, but right now doctors track them all, said study author Christopher Wolfgang, director of surgical oncology at the Kimmel Cancer Center. “We need to follow all patients, on the order of hundreds of thousands of patients, with expensive and, in some cases, invasive tests to find those few patients who will progress to cancer.” Follow-up testing can involve radiation exposure and complications, as well as provoke anxiety, he adds.
Lennon, Wolfgang, and others set out to build a tool to sift through patient information in the hopes of identifying patterns to distinguish low-risk from high-risk cysts. To do so, they gathered data from hundreds of patients at Hopkins and 15 medical centers around the world who were diagnosed with a cyst and then underwent surgery to have it removed. After surgery, each cyst was examined and classified as having either no risk, a small risk, or a high risk of progressing to cancer.
The team’s test, CompCyst, is focused around a machine learning algorithm called MOCA, for Multivariate Organization of Combinatorial Alterations, that combines molecular data—including DNA mutations and chromosomal changes—with protein information from extracted cyst fluid and imaging tests.
The team trained the algorithm with data from 436 patients, then tested it on a second, separate set of data from 426 patients. The algorithm tests millions of combinations of the data points to predict the right treatment pathway with high sensitivity and specificity, said co-author Marco Dal Molin, a postdoctoral research fellow at Hopkins.
CompCyst outperformed the standard-of-care that doctors use today in all three patient groups: The test correctly predicted 60 percent of patients who should have been sent home (versus 19 percent using standard-of-care), 49 percent of patients who should have been monitored (versus 34 percent), and 91 percent of patients in need of surgery (versus 89 percent).
Overall, the researchers estimate that if CompCyst had been used to decide care for these patients, 60 to 74 percent of them (depending on the type of cyst) would have avoided unnecessary surgery.
Combining clinical and genetic features using machine learning is “the wave of the future to inform clinical judgment not only about pancreatic cysts but about many other diseases,” said co-author Bert Vogelstein, a professor of oncology and co-director of the Ludwig Center at Hopkins.
Vogelstein and two additional co-authors recently co-founded a company, Thrive Earlier Detection, that has licensed CompCyst for commercial development.
Megan is an award-winning freelance journalist based in Boston, Massachusetts, specializing in the life sciences and biotechnology. She was previously a health columnist for the Boston Globe and has contributed to Newsweek, Scientific American, and Nature, among others. She is the co-author of a college biology textbook, “Biology Now,” published by W.W. Norton. Megan received an M.S. from the Graduate Program in Science Writing at the Massachusetts Institute of Technology, a B.A. at Boston College, and worked as an educator at the Museum of Science, Boston.