Researchers' AI Predicted Pancreatic Cancer 3 Years Before Doctors

The Register reports: AI algorithms can screen for pancreatic cancer and predict whether patients will develop the disease up to three years before a human doctor can make the same diagnosis, according to research published in Nature on Monday.

Pancreatic cancer is deadly; the five-year survival rate averages 12 percent. Academics working in Denmark and the US believe AI could help clinicians by detecting pancreatic cancer at earlier stages, if the software can reliably predict which patients are at higher risk of developing the disease. The researchers trained AI algorithms on millions of medical records obtained in the Danish National Patient Registry and the US Veterans Affairs Corporate Data Warehouse... "Cancer gradually develops in the human body, often over many years and fairly slowly, until the disease takes hold," Chris Sander, the study's co-senior investigator and leader of a lab working at the Department of Systems Biology at Harvard Medical School, told The Register. "The AI system attempts to learn from signs in the human body that may relate to such gradual changes..."

"AI on real-world clinical records has the potential to produce a scalable workflow for early detection of cancer in the community, to shift focus from treatment of late-stage to early-stage cancer, to improve the quality of life of patients and to increase the benefit/cost ratio of cancer care," the paper reads... The study is still in its early stages, and the software cannot yet be used to run screening programs. Improvements are needed before even a trial can be conducted... Still, the team believes that as the technology improves and operating costs decrease, AI could become a valuable screening tool in the future. "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," said Søren Brunak, professor of disease systems biology and director of research at the Novo Nordisk Foundation Center for Protein Research at the University of Copenhagen, a co-senior investigator of the study, said in a statement.

"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," he concluded.
Thanks to Slashdot reader Tony Hu for sharing the article.

Are they cherry picking data?

[wakeboarder]

Now apply this AI test to everybody and see how well it is able to predict cancer in everyone not just one case

Re:

[XXongo]

Yep.

They're not cherry-picking data because they don't have any data to cherry-pick.

It hasn't predicted any cancers; it's run on historical data only. From the summary: "Improvements are needed before even a trial can be conducted... Still, the team believes that as the technology improves and operating costs decrease, AI could become a valuable screening tool in the future..."

Re:

[phantomfive]

This kind of data science (which is called AI these days because that gets funding) is pretty common in the medical field. Find a correlation, then hope it works out. Lots of people are working to find a way to use hospital data to save lives (mainly because it would be profitable to do so). A few years ago there were a bunch of startups trying to use hospital data to save people from sepsis.

Re:

[Fons_de_spons]

Fals positivity rate would be very interesting in that case...
Saw a statistical breakdown of the Theranos tests once. They tested so much that the chance that one of the tests came out falsly positive would be pretty big. (Even assuming that everything was done by the book).

Re:

[test321]

I think false negatives matter much more. There is a huge liability risk in telling someone they're fine and 3 years later they get most deadly cancer known. Even a larger rate of false positive still narrows down quite a lot the population that need to be included into a moderately priced periodic surveillance (echography), and reduces a lot the cost for whoever pays (whether it's a private insurance or the public social security) as compared to: surgery, chemotherapy, palliative care, followed by 88% deat

Re:

[madsh]

Well there is another old piece in Nature titled ‘When the nation is the cohort’. So cherry picking is not my first question. I think Brunak is on to something here. And I think we need to talk about what we do with the growing patient records. When should research and governments begin to notify people about predictions about diseases?

Re:

[dbialac]

Novel if initially creepy sounding question: Why exert all of these efforts trying to fix organs that go wrong? Why not just completely replace the internal organs with a mechanical system that can handle all of the necessary tasks (blood flow, oxygenation, etc.) that the rest of our bodies need with mechanical systems that don't get cancer, etc.?

Re:

[wakeboarder]

They already do this, but those things need batteries. Why would you want to charge yourself to stay alive.

Re:

[dbialac]

The entire body cavity? Regardless, doesn't sleep do effectively the same thing?

Re:

[robi5]

By the time some of the cancer types, like pancreatic cancer, are typically detected, they have often spread to other organs. While all organs are hard or currently impossible to replace, remember that the brain is also an organ, the skin is also an organ, the spine is an organ, there are bones, face, senses, vocal chords, limbs.

The title dosn't seem accurate

[strike6]

The article doesn't say they have detected Pancreatic Cancer three years earlier than a doctor. They are predicting that they can. It doesn't sound like these algorithms have been used clinically at all. And according to the article, the algorithms aren't even good enough to be used for any type of real world clinical trials yet. This seems very premature to be talking about. Get back to me when there is a viable test run/clinical trial.

Re:

[quonset]

Get back to me when there is a viable test run/clinical trial.

Yes, yes. It's not 100% accurate or predictive, so we can dismiss it out of hand. Can't use it for anything useful such as testing patients who might be predicted to get pancreatic cancer. Can't use it in any way to get a jump on treatment. It's completely useless.

It's like seat belts. They're not 100% effective so nobody should use them.

Re:

[strike6]

Way to cherry pick parts of my post. I never said they shouldn't pursue this. I said: "This seems very premature to be talking about. Get back to me when there is a viable test run/clinical trial." I stand by that.

Re:

[timeOday]

It's not that. It wouldn't have been published in Nature if it weren't promising, and I assume they didn't do anything really silly like publishing the fit to the training data as predictive results. But if you've had access to the test data during model development, it's hard to 100% eliminate data leakage. It's just not the same until you have taken 1 model frozen in time and then used it on data that was created afterwards. In fairness the article does contain multiple caveats.

It also points out the

Insurance companies

[gillbates]

The accuracy is much more important to the individual than to the insurance company. If a test predicted cancer with 50% accuracy, patients would want a better test. However, if an insurance company could drop an individual with a 50% chance of cancer three years before they actually developed cancer, it could dramatically increase their profit margin. From the insurance company's standpoint, a test which could predict future adverse events beyond the statute of limitations would be most beneficial.

- big medicine can bring improvements -

[swell]

There seems to be potential in this approach. Note the comment: "Effective prediction in real-world settings will rely on the quality of patients' medical histories."--my (US Veteran's Administration) medical history would require hundreds of pages to print out. More standardization of such files across countries and institutions is desperately needed.

Only the Nature link https://www.nature.com/article... [nature.com] has the facts of the process. Quite tedious. Vast numbers of authors from many institutions were involved. They expended great effort only to show that a better study is needed. It's too late for my pancreas cancer, but others will eventually benefit.

False positives

[groobly]

A test can predict pancreatic cancer 100% of the time ahead of the time that "doctors" find it, as long as it predicts that everyone will get pancreatic cancer.

These folks need to show us some kind of ROC before yapping.

Re:

[ewibble]

Doctors can't either, the question is which is more accurate and will it save lives and be affordable. Asking for for 100% is just ridiculous. Just like self driving cars the question isn't are they 100% safe (of course not). The question is are they safer than people.

I also assume they took into account accuracy including false positives, unless they are complete morons.

false positives not all that worrisome here

[david.emery]

Pancreatic cancer isn't all that common. Now if the prediction algorithm flags a lot of false positives and that generates more careful/more frequent screening, that's Not A Bad Thing. Now I'm sure the insurance companies wouldn't be happy at, for instance, 3x false positives causing them more expense for "unneeded" screening, but (as a diabetic who is more likely to get this,) I'd be happy for any advance warning of this disease, even if it's a false positive.

Re:false positives not all that worrisome here

[jsonn]

You haven't had to deal with a cancer diagnosis, have you? There is a reason why the intervals for mammographic screening have been significantly reduced for patients under 40 since the 1980s: the false positive rate and resulting stress and harm far outweighed the gains.

Re:

[dbialac]

I'm a cancer survivor. It was only by chance that my cancer was caught early enough to likely have been surgically removed permanently. An early screening for another condition caught it and it otherwise wouldn't have been caught until much later. As for me, I still need to be screened every 6 months to verify that it's still gone.

Re:

[jsonn]

I want to stress that the screenings in your case are done for a completely different reason and likely also using a different methodology. I would illustrate the difference by giving urine-based point-of-care pregnancy tests to (female) 12-year-olds. Google suggests a specificity of 99.2% for those tests. A random female between 15-19 in the USA has a chance of 1.6% to have been pregnant, so I would strongly hope that the chance for a 12-year-old is significantly lower. At this point, the chance that the p

AI or data mining?

[trevc]

Are things like this really AI or just data mining?