AI can predict pancreatic cancer three YEARS before it occurs, major Harvard study finds
Using medical records and information from previous scans, the AI was able to flag patients at a high risk of developing pancreatic cancer within the next three years with great accuracy.
There are currently no full-proof scans for pancreatic cancer, with doctors using a combination of CT scans, MRIs and other invasive procedures to diagnose it. This keeps many doctors away from recommending these screenings.
The study has doctors hopeful because pancreatic cancer is notoriously hard to spot, making it one of the deadliest forms of the disease, killing more than half of sufferers within five years of diagnosis.
Over time, they also hope these AI models will help them develop a reliable way to screen for pancreatic cancer — which already exists for other types of the disease.
Unlike other cancers, there is no single way to screen for it and, in the early stages, it can cause mild symptoms that are often overlooked.
'One of the most important decisions clinicians face day to day is who is at high risk for a disease, and who would benefit from further testing, which can also mean more invasive and more expensive procedures that carry their own risks,' Dr Chris Sander, a biologist at Harvard who contributed to the study, said.
'An AI tool that can zero in on those at highest risk for pancreatic cancer who stand to benefit most from further tests could go a long way toward improving clinical decision-making.'
The National Cancer Institute estimates 64,050 Americans will be diagnosed with pancreatic cancer this year, and it will be responsible for 50,550 deaths.
The American Society of Clinical Oncology estimates that 56 percent of all people diagnosed with die from the disease.
If the cancer spreads to another part of the body — called metastasis — the survival rate falls to just three percent.
This makes finding a way to screen for pancreatic cancer early crucial, as any delay in treatment greatly increases a person's risk of death.
'Many types of cancer, especially those hard to identify and treat early, exert a disproportionate toll on patients, families and the healthcare system as a whole,' Dr Soren Brunak, a Danish study author, said.
'AI-based screening is an opportunity to alter the trajectory of pancreatic cancer, an aggressive disease that is notoriously hard to diagnose early and treat promptly when the chances for success are highest.'
Harvard researchers partnered with scientists at Danish pharma giant Novo Nordisk, among other from the US and Denmark for their study, published Monday in Nature Medicine.
They trained their AI model by using 500 CT scans of people who had experienced lung nodules.
These are abnormal growths within the lungs. They are not cancerous in 95 percent of cases.
However, they can also serve as a sign that pancreatic cancer has spread to the lungs.
Then, they used past medical records to see if the AI could accurately identify people who were more likely to be diagnosed with the disease.
In total, data from 6million Danes and 3million Americans were included. Among the study population, 24,000 people from Denmark suffered pancreatic cancer, along with 3,900 Americans.
The 9million data points were fed into the AI, and it was tasked with predicting a person's likelihood of suffering pancreatic cancer within the next three years.
They gauged their model's accuracy by generating an 'area under the curve', or AUC, score.
These scores work by comparing results of the model to the real life patient outcomes. It generates a score between 0 and 1.0.
A model that receives a 0 is worthless, 0.5 is as accurate as flipping a coin, and 1.0 indicates a perfect model.
Generally, scientists consider a score of 0.8 or higher to be an indication of an accurate test.
The Harvard model earned a score of 0.88 in estimating risk of cancer within the next three years, and 0.9 for detecting risk in the next 12 months.
It also was tested to see if it would predict further intervention for people with scans that would be considered a 'medium risk' of developing cancer.
Among 22 people with lung nodules that eventually were diagnosed with the cancer, the AI flagged 18 as having a high risk of developing the disease.
Researchers hope this model can be used by doctors in cancer care. Current screening tools include an MRI, CT scans or endoscopic ultrasounds — where a doctor inserts a camera deep down a person's throat.
These can be uncomfortable, resource intensive and expensive, though. This makes doctors hesitant to recommend them to patients.
None are 100 percent accurate ways to find the cancer either.
However, using this type of tool to flag high-risk patients can help doctors make sure the people who need these screenings the most can get them.