Slashdot Mirror
AI Researchers Predict Alzheimer's Years Before Diagnosis (sciencedaily.com)
Slashdot reader pgmrdlm quotes Science Daily: Timely diagnosis of Alzheimer's disease is extremely important, as treatments and interventions are more effective early in the course of the disease. However, early diagnosis has proven to be challenging. Research has linked the disease process to changes in metabolism, as shown by glucose uptake in certain regions of the brain, but these changes can be difficult to recognize... Researchers trained [a] deep learning algorithm on a special imaging technology known as 18-F-fluorodeoxyglucose positron emission tomography (FDG-PET). In an FDG-PET scan, FDG, a radioactive glucose compound, is injected into the blood. PET scans can then measure the uptake of FDG in brain cells, an indicator of metabolic activity.
The researchers had access to data from the Alzheimer's Disease Neuroimaging Initiative (ADNI), a major multi-site study focused on clinical trials to improve prevention and treatment of this disease. The ADNI dataset included more than 2,100 FDG-PET brain images from 1,002 patients. Researchers trained the deep learning algorithm on 90 percent of the dataset and then tested it on the remaining 10 percent of the dataset. Through deep learning, the algorithm was able to teach itself metabolic patterns that corresponded to Alzheimer's disease. Finally, the researchers tested the algorithm on an independent set of 40 imaging exams from 40 patients that it had never studied. The algorithm achieved 100 percent sensitivity at detecting the disease an average of more than six years prior to the final diagnosis.
"We were very pleased with the algorithm's performance," Dr. Sohn said. "It was able to predict every single case that advanced to Alzheimer's disease
The researchers had access to data from the Alzheimer's Disease Neuroimaging Initiative (ADNI), a major multi-site study focused on clinical trials to improve prevention and treatment of this disease. The ADNI dataset included more than 2,100 FDG-PET brain images from 1,002 patients. Researchers trained the deep learning algorithm on 90 percent of the dataset and then tested it on the remaining 10 percent of the dataset. Through deep learning, the algorithm was able to teach itself metabolic patterns that corresponded to Alzheimer's disease. Finally, the researchers tested the algorithm on an independent set of 40 imaging exams from 40 patients that it had never studied. The algorithm achieved 100 percent sensitivity at detecting the disease an average of more than six years prior to the final diagnosis.
"We were very pleased with the algorithm's performance," Dr. Sohn said. "It was able to predict every single case that advanced to Alzheimer's disease
What do they tell the patient? "Good news! You won't start losing your marbles until six years from now?"
Everyone get tested, so that people who come down with Alzheimer's will be called a pre-existing condition.
The shepherds did so well protecting the flock that the sheep no longer believed that wolves existed.
IF it's really 100% (or close) accurate, how soon can it be used for the general populace?
Alzheimer's costs a LOT to deal with. Beyond avoiding some suckage for the patient and especially their family, this could save some serious money if universally applied.
It's kinda invasive with requiring an injected marker, but that's already done pretty routinely for kidney trouble etc.
The algorithm achieved 100 percent sensitivity at detecting the disease an average of more than six years prior to the final diagnosis.
So no false negatives. But how many false positives? TFA doesn't say.
You can get 100% detection of fires by sounding the alarm continuously. But that's not very useful. Without that other number we don't know if this is useful or bull generated fertilizer.
Bantam Dominique roosters crow a four-note song. Once you've heard it as "Happy BIRTHday" you can't NOT hear it that way
Comment removed based on user account deletion
If you want to diagnose if a protein gets made, take a sample of it, put it in a time machine to speed up time, see if it develops the protein. Perfect detection, 100% success rate.
It should be quite easy.
i.e. if all matter and lights interactions are dipolar, rate of time depends on dipolar spin rate, if you speed up the spin rate, you speed up time.
Postulate G: https://hardware.slashdot.org/comments.pl?sid=12877158&cid=57615906
Or if you want to wait for a cure, climb in the time machine and slow it down.
Dipolar spin is an electric spin, so you need some fancy induction oven. Universes spin frequency is F, so it needs to induce a spin F*0.99, then F*0.98.... slowing down the constant.
In the process of slowing or speeding up time, you'll induce a second dipolar spin component. Which applies a force, which breaks up particles into photons. So you want to slow that number down real slow.
Honestly I wanted to make a Warp Drive or Faster Than Light computer as proof of dipole theory, but meh. I guess a time machine will have to do.
Man there are so few physics stories on Slashdot.
I want to talk about what an anti-particle is, about why particles are prime numbers multiples of the spin frequency, and how it cancels matter, then light bending, then the speed of gravity over space.
And my finally was to make a Warp drive, but. Meh.
Maybe there will some future viruses, that can infect the neural network? And the machines will just be screaming, trying to die. But he cannot die.
Just screaming forever..screaming and dying forever.
But it forgot to tell us.
How much Does changing your diet slow it down? A few percentage points is hardly worth it. The fact is there still is no cure and the latest research is finding strong ties with viral infections of the brain, especially by herpes simplex one (cold sore virus)
for a loooong time already. Also do not buy this vendor lock in pseudo security, not built-into hard-drives, or this Apple T2 chip thing. Can't review, don't know what it does, can't trust. Use open source reviewable stuff that we can also fix in case of an issue. I for one only trust open solutions an non of the pseudo "military grade" fluff that breaks as soon as someone looks too close, https://www.youtube.com/watch?...
Talk about sensitivity (how few false negatives) is important, but you also need high specificity - i.e. how few false positives.
In this case, their specificity was only 82% of the time, that is, a significant issue.
Human doctors were less likely in general to say you had the disease, with only a sensitivity of about 57% but their greater reluctance to say "You are mentally impaired) meant they had a 91% specificity.
Personally, I would not go to a doctor that told one in five healthy people they were sick.
Look at real numbers;
200 people. 1 in eight have the disease, which works out to 25 people. The computer would tell those 25 people they have the disease AND also mistakenly tell 150*(100-81=18%), or about 27 healthy people they have the disease when they do not have it. More people would be falsely told they are sick than are actually sick.
The machines are not really good enough to be used yet.
For humans, they would catch about 14 of the 25 sick people, but would have told 17 healthy people told they were sick. Given that treatment is very limited, better to stick with the human doctor, rather than scaring 10 healthy people in order to help an extra 11 people prepare for something they can't do much about.
excitingthingstodo.blogspot.com
Related situation : Look up "predictive autoimmunity". /s And I'm sure these techniques will be available at Kaiser Permanente in about 40-50 years ;--) s/
If I understand it correctly, assaying a variety of enzymes will predict various auto-immune conditions 7-10 years before they are currently detectable. I suspect that similar tests can predict or detect MANY other conditions and diseases way sooner than current techniques.
Many conditions have MUCH better ways of being detected, treated, and cured than the current "standard of care". Think how many people with vestibular issues, tinnitus, or post-concussion issues, could benefit from neuroplastic healing if it were more widely available. As in all things - follow the money.
It didn't 'teach itself'. It was fed data and carried out the operations on it it was programmed to carry out. At least the doctors had the decency to call it an algorithm and not OMG AI.
It's very hard to rule out that a patient might not get alzheimer's in the future and/or may be on the path to developing alzheimer's but then die before it becomes symptomatic. So it's much harder to define a set of high confidence "negative" pet scans for alzheimers to evaluate the accuracy of the network than to develop the set of high confidence positive pet scans.