Skip to main content

Machine Learning Guides Pancreatic Cyst Management in Patients

pancreatic cyst assessment
Comparing recommendations for managing patients with pancreatic cysts based on either CompCyst or standard-of-care pathology. | S. Springer et al., Science Translational Medicine (2019)]

Researchers have created a comprehensive test based on machine learning algorithms to better guide the management of patients with pancreatic cysts — a potential precursor of pancreatic cancer. The new assay called CompCyst was tested in an international, multicenter study of 875 patients, where it could reduce the number of unnecessary surgeries for benign cysts by 75%. The research was published in the July 17 issue of Science Translational Medicine.

The assay could allow clinicians to more definitively distinguish between precancerous and benign pancreatic cysts, potentially leading to fewer unneeded surgeries and lower health and economic costs. The authors highlight that future work will be necessary to prospectively confirm the markers used in the test, but say their platform has strong clinical potential as a complement to existing approaches.

"When we incorporate all these markers, we find we can do a much better job of determining the cyst type that the patient has," said Bert Vogelstein, Clayton professor of oncology, Howard Hughes Medical Institute investigator and co-director of the Ludwig Center at the Johns Hopkins Kimmel Cancer Center and a senior author of the new study.

"We can reduce the number of surgeries if we use this test by 75%, which means that many, many patients who would undergo needless surgery would not have to," he added.

Pancreatic cysts are fluid-filled lesions in the pancreas that are found in up to 8% of all people over the age of 70, according to a 2010 study. Although most cysts are benign, some cysts that produce mucin — the chief component of mucus — can transform into a dangerous form of pancreatic cancer named pancreatic ductal adenocarcinoma (or PDAC). PDAC is the third leading cause of cancer deaths and often has dismal survival rates, usually because the tumor is identified in its advanced stages.

Catching pancreatic cancer early remains an ongoing goal for clinicians, but accurately diagnosing pancreatic cysts can be problematic, according to the study authors. First, it is often difficult to distinguish mucin-producing, precancerous cysts from benign ones using pathological analysis of the removed cyst, which is considered the current gold standard of diagnosis. Second, it is challenging to determine whether mucin-producing cysts are indicative of advanced or early disease.

As a result, noncancerous cysts are often misclassified and unnecessarily removed with pancreatic surgery — a procedure that poses major health risks and can lead to death in up to 5% of cases. For example, one study estimated that 25% of patients who underwent surgery for a pancreatic cyst had a benign cyst with no cancerous potential.

The inability to differentiate between high and low risk lesions also means clinicians must follow hundreds of thousands of patients with expensive imaging and fluid tests over many years to find those few patients who will progress to cancer, said Chris Wolfgang, an associate professor of surgery at the Johns Hopkins University School of Medicine and a senior author of the study.

Monitoring so many people places a major strain on healthcare resources and exposes many patients to radiation and other health risks associated with these tools, he said.

"The purpose of our study was to try to improve the diagnosis of these cysts in patients who have them, and try to predict which patients actually need surgery, and which patients don't," said Vogelstein.

To overcome the dilemma faced by clinicians, Simeon Springer, a scientist at Ring Therapeutics and colleagues enrolled the 875 patients with pancreatic cysts that were surgically removed. They comprehensively analyzed the genetic and molecular profiles of the cysts and collected information about the mutations, proteins and other markers linked to either benign or mucin-producing cysts.

The authors then combined these markers with imaging data and established machine learning algorithms to create CompCyst. They used deep learning techniques to train the platform to read these markers and classify patients into those who should be monitored, not monitored, or receive surgery, based on the likelihood of their cyst being precancerous.

After training the test protocol with 436 of the original patients, the researchers validated CompCyst by testing it in 426 other patients. They found the test largely outperformed standard-of-care pathology: it correctly identified 60.4% of patients who should have been discharged (versus 18.9% using standard-of-care diagnosis), 48.6% of patients who should have been monitored (versus 34.3%), and 90.8% of patients in need of surgery (versus 88.8%).

Applying CompCyst would have avoided surgery in 60% of the 193 patients who underwent unnecessary surgical removal for their cysts, according to the authors.

"Our study directly addresses the limitations of current clinical criteria and has the potential to be a major leap forward in the management of pancreatic cysts," said Wolfgang. "If these results translate into clinical practice, a large number of patients will be spared an unnecessary operation with associated mortality and lifelong morbidity."

The test still faces various limitations before it can be applied to the clinic, as the cysts analyzed were not representative of those seen in routine practice. The team also stressed that CompCyst is not meant to replace conventional tools, but rather to provide clinicians with greater confidence when advising and treating patients.

"The test has been developed, but now it has to be applied to many more patients who have not yet undergone surgery," said Vogelstein. "Over the course of the next five years, we expect to use CompCyst to study thousands of patients who have cysts, follow them along, advise surgery or no surgery when needed, and determine how well the test does in the real world."

The team estimates that the test will be available within six to 12 months at Johns Hopkins, according to Vogelstein. In the long-term, they hope to carry out a new prospective study that will lead to FDA approval of the test, which could then be commercialized and offered to the public.

[Credit for related image of pancreatic adenocarcinoma: Ed Uthman/ Flickr]


Joseph Cariz

Related Focus Areas