AI Tool Predicts Risk of Lung Cancer

‘Sybil’ accurately predicted individuals’ lung cancer risk 1 to 6 years into the future

By HALEY BRIDGER | MGB Communications Research Care Delivery

5 min read

Illustration of lungs with graphs & charts
Image: greenbutterfly/iStock/Getty Images Plus

Lung cancer is the leading cause of cancer death in the United States and around the world and low-dose chest computed tomography (LDCT) is recommended to screen people between 50 and 80 years old who have a significant history of smoking or currently smoke.

Lung cancer screening with LDCT has been shown to reduce deaths from lung cancer by up to 24 percent, but as rates of lung cancer climb among nonsmokers, new strategies are needed to screen and accurately predict lung cancer risk across a wider population.

Get more HMS news here

A study led by Harvard Medical School investigators at Massachusetts General Hospital, in collaboration with researchers at MIT, developed and tested an artificial intelligence tool known as Sybil.

Related Stories

Heart-Saving AI

An artificial-intelligence system shows promise in identifying signs of heart transplant rejection.

Illustration of a heart x-ray with technology around it, such as a CD

Based on analyses of LDCT scans from patients in the U.S. and Taiwan, Sybil accurately predicted the risk of lung cancer for individuals with or without a significant smoking history.

Results are published in the Journal of Clinical Oncology.

“Lung cancer rates continue to rise among people who have never smoked or who haven’t smoked in years, suggesting that there are many risk factors contributing to lung cancer risk, some of which are currently unknown,” said corresponding author Lecia Sequist, the HMS Landry Family Professor of Medicine in the Field of Medical Oncology at Mass General.

“Instead of assessing individual environmental or genetic risk factors, we’ve developed a tool that can use images to look at collective biology and make predictions about cancer risk,” said Sequist, who is also director of the Center for Innovation in Early Cancer Detection at Mass General and a medical oncologist specializing in lung cancer.

The U.S. Preventive Service Task Force recommends annual LDCTs for people over age 50 with a 20 pack-year history of smoking, who either currently smoke or have quit smoking within the last 15 years. But less than 10 percent of eligible patients are screened annually.

To help improve the efficiency of lung cancer screening and provide individualized assessments, Sequist and colleagues at the Mass General Cancer Center teamed up with investigators from the Jameel Clinic at MIT.

Science. Medicine. Humanity.

Harvard Medicine magazine delivered to your inbox

Subscribe to our free monthly e-letter

Using data from the National Lung Screening Trial (NLST), the team developed Sybil, a deep-learning model that analyzes scans and predicts lung cancer risk for the next one to six years.

“Sybil requires only one LDCT and does not depend on clinical data or radiologist annotations,” said co-author Florian Fintelmann, HMS associate professor of radiology at Mass General.

“It was designed to run in real time in the background of a standard radiology reading station, which enables point-of care clinical decision support,” he said.

Independent data sets

The team validated Sybil by using three independent data sets — a set of scans from more than 6,000 NLST participants who Sybil had not previously seen, 8,821 LDCTs from Mass General, and 12,280 LDCTs from Chang Gung Memorial Hospital in Taiwan.

The latter set of scans included people with a range of smoking histories, including those who never smoked.

Sybil was able to accurately predict risk of lung cancer across these sets. The researchers determined Sybil’s accuracy by using area under the curve (AUC), a measure of how well a test can distinguish between disease and normal samples and in which 1.0 is a perfect score.

Sybil predicted cancer within one year with AUCs of 0.92 for the additional NLST participants, 0.86 for the Mass General data set, and 0.94 for the data set from Taiwan.

The program predicted lung cancer within six years with AUCs of 0.75, 0.81, and 0.80, respectively, for the three data sets.

“Sybil can look at an image and predict the risk of a patient developing lung cancer within six years,” said co-author and Jameel Clinic faculty lead Regina Barzilay, a member of the Koch Institute for Integrative Cancer Research.

“I am excited about translational efforts led by the MGH team that are aiming to change outcomes for patients who would otherwise develop advanced disease,” Barzilay said.

The researchers noted that this is a retrospective study, and prospective studies that follow patients going forward are needed to validate Sybil.

In addition, the U.S. participants in the study were overwhelmingly white (92 percent), and future studies will be needed to determine whether Sybil can accurately predict lung cancer among other populations.

Sequist and colleagues will be opening a prospective clinical trial to test Sybil in the real world and understand how it complements the work of radiologists. The code has also been made publicly available.

“In our study, Sybil was able to detect patterns of risk from the LDCT that were not visible to the human eye,” said Sequist. “We’re excited to further test this program to see if it can add information that helps radiologists with diagnostics and sets us on a path to personalize screening for patients.”

Authorship, funding, disclosures

This work was supported by the Bridge Project, a partnership between the Koch Institute for Integrative Cancer Research at MIT and the Dana-Farber/Harvard Cancer Center, as well as the MIT Jameel Clinic, Quanta Computing, Stand Up To Cancer, and the Massachusetts General Hospital Center for Innovation in Early Cancer Detection, and includes support from the Bralower and Landry Families and Upstage Lung Cancer. Funding to support this research was also provided for by the Eric and Wendy Schmidt Center at the Broad Institute of MIT and Harvard.

Disclosures provided by the authors are available with this article at https://doi.org/10.1200/JCO.22.01345.