Pancreatic diagnoses have been on the rise over the past several years—maybe a decade or so—and that is largely because it has been overlooked for so long. Indeed, pancreatic cancer patients often die from the condition because it is diagnosed so late in the disease progression that very little, if anything, can be done, and the tumors have spread beyond help. Other patients still die after the removal of benign cysts that tend to appear threatening amid a handful of other confusing data.
Still, doctors at Johns Hopkins have used artificial intelligence to develop a system that can provide a clearer picture of pancreatic cancer in patients. When testing for the condition, the research team found it provided an immensely remarkable ability to differentiate which lesions are harmful and which pose no threat to the patient at all. This, of course, will help reduce unnecessary—and, apparently, life-threatening—surgeries.
Experts estimate 800,000 patients are diagnosed with pancreatic cysts in the United States every year. Unfortunately, doctors have long lamented there is no good way to tell which cysts harbor one of the most deadly forms of cancer and which cysts are benign. This ambiguity leads to several thousand unnecessary surgeries every year. In fact, one study determined that as much as 78 percent of surgeries recommended for removing potentially cancerous pancreatic cysts are completely unnecessary. Apparently the vast majority of pancreatic cysts are, in fact, benign, but doctors track them all the same. This new method, then, not only reduces risk for unnecessary surgery, but can cut the associated costs as well.
And this is why the new machine learning algorithm is so important. This new technique, called CompCyst (comprehensive cyst analysis), is better than today’s standard of care in terms of predicting pancreatic cancer. To develop this method, scientists gathered data from several hundred patients at more than a dozen medical centers around the world, who had cysts removed. Each cyst was examined and classified for risk.
The test uses a machine learning algorithm called Multivariate Organization of Combinatorial Alterations, or MOCA. It combines molecular data (DNA mutations, chromosomal changes, etc) with protein information taken from extracted cyst fluid, as well as imaging tests.
In all, the test correctly identified patients who should have been sent home (versus less than 20 percent using standard-of-care), 49 percent who should have been monitored (compared with 34 percent) and 91 percent of patients who should have been diagnosed for surgery (on par with standard of care).
The study has been published in the journal Science Translational Medicine.