AI Can Detect Alzheimer’s Disease Six Years Before a Diagnosis

https://www.ucsf.edu/news/2018/12/412946/artificial-intelligence-can-detect-alzheimers-disease-brain-scans-six-years

by laurex • 8 years ago

259 101 8 years ago

classics2• 8 years ago

Typically Alzheimer’s diagnoses are given with extreme reluctance, it changes how every person that interacts with that patient will view their condition, prognosis and treatment options.

I think that has more to do with the extended time frame than the actual ability of family and medical providers to see changes in a patient cognition and abilities.

djohnston• 8 years ago

i think these early diagnostic techniques are a necessary brick on the road to curing AD. by the time people are showing clinical symptoms the disease has already progressed and has been doing so for decades. treatments that slow progression will be more effective if administered earlier on in the progression.

rincebrain3• 8 years ago

That seems like a problem subject to heavy confirmation bias - e.g. people decide their relative is cognitively impaired and see any transient shortfall as being more evidence, rather than any number of other possible causes.

(Not that you can't necessarily see cognitive decline externally earlier than a diagnosis, just that it seems like it'd be hard to make a clinical argument about persistent decline versus several poor nights of sleep, for example.)

vegardx9• 8 years ago

It also triggers a lot of legal implications. In Norway your divorce is automatically approved if your spouse is diagnosed with Alzheimer's.

tigershark• 8 years ago

Seriously? It seems really inhuman to me.

dsfyu404ed2• 8 years ago

Seems like a great incentive to pay off a doctor to diagnose someone so you can lose less money, get the kids, have another bargaining chip in a messy divorce, etc, etc.

Edit: Pointing out flaws/possible exploits in a system is not the same thing as endorsing their use.

aaaaaaaaaab• 8 years ago

Approved? Why is an approval even needed? They can force people to stay married in Norway by not approving a divorce?

pvaldes3• 8 years ago

And this people can be also sort of 'legally sequestered' by the government. I know a case and is horrifying to see it.

offsetr• 8 years ago

Some people with a unique gift in sense of smell can do it even better than 6 years.

socceroos• 8 years ago

Anecdotally, and slightly off-topic, I was able to smell when my wife was pregnant before we were able to find out with a test or late period. Not sure if that's relatively normal or not though.

londons_explore• 8 years ago

That "not a sexy person" smell in the early stages of pregnancy must kill a lot of relationships at just the wrong time...

sierdolij• 8 years ago

Sorry, what do you mean? Are there examples?

rincebrain3• 8 years ago

There was a specific case of a woman who realized, after her husband acquired a peculiar smell and was 6 years later diagnosed with Parkinson's, then everyone at a Parkinson's support group had that smell, that she could smell Parkinson's in people well before any clinical tests.

She went to researchers who tested her with sweat from 6 diagnosed patients and 6 controls - she identified the 6 diagnosed patients correctly, and one of the controls who was later diagnosed with it but had not been yet.

https://www.telegraph.co.uk/science/2017/12/18/woman-can-sme...

(Also, as a fascinating aside which probably isn't what the GP comment meant, _loss_ of sense of smell is one of the earliest symptoms of Alzheimer's or Parkinson's. Why? Good question.)

seanhunter• 8 years ago

There's a reasonable amount of published literature on this type of thing. For example in the case of Parkinsons https://news.nationalgeographic.com/2018/01/smell-sickness-p...

jobigoud• 8 years ago

No cure exists for Alzheimer’s disease, but promising drugs have emerged in recent years that can help stem the condition’s progression.

What are the downsides of these treatments on a healthy brain? Why are we not administering them preventively to anyone with a family history of AD?

ralfd• 8 years ago

promising drugs have emerged in recent years

To recent yet.

vegardx9• 8 years ago

We are, sort of. We've added lithium to drinking water which has been linked to decrease in dementia.

People that suffer from bipolar disorder and are being treated with lithium seems rarely if ever to develop dementia, but the dosages are much much (1000 times, or so) higher than what we add to drinking water.

According to my doctor the primary issue with adding higher concentration of lithium (still, way below therapeutic levels for mental disorders) to drinking water is quacks that make a lot of noise when the topic comes up. They see it as adding mind altering chemicals to water, which couldn't be further from truth.

moreira• 8 years ago

Well, if they’re reducing the incidence of dementia, and you would’ve gotten dementia otherwise... they -did- alter your mind. And that’s a good thing.

People and their fear of chemicals though. Reminds me of http://www.dhmo.org

quotemstr• 8 years ago

Lithium is linked to all sorts of behavioral changes though. Who knows what microdosing the entire population would do? I really don't think it's fair to describe opposition to this sort of project as arising from "quacks".

tluyben2• 8 years ago

But why not prevently dose families with high occurrence far higher instead of the entire population?

logicchains2• 8 years ago

According to https://www.webmd.com/vitamins/ai/ingredientmono-1065/lithiu... (presumably about as mainstream a source as one can hope to find online):

"Lithium can cause nausea, diarrhea, dizziness, muscle weakness, fatigue, and a dazed feeling. These unwanted side effects often improve with continued use. Fine tremor, frequent urination, and thirst can occur and may persist with continued use. Weight gain and swelling from excess fluid can also occur. Lithium can also cause or make skin disorders such as acne, psoriasis, and rashes worse. The amount of lithium in the body must be carefully controlled and is checked by blood tests."

"Lithium can poison a developing baby (fetus) and can increase the risk of birth defects, including heart problems.

Lithium treatment is UNSAFE in women who are breast-feeding. Lithium can enter breast milk and cause unwanted side effects in a nursing infant.

Heart disease: Lithium may cause irregular heart rhythms. This may be a problem, especially for people who have heart disease.

Kidney disease: Lithium is removed from the body by the kidneys. In people with kidney disease, the amount of lithium that is given might need to be reduced.

Surgery: Lithium might change levels of serotonin, a chemical that affects the central nervous system. There is some concern that lithium might interfere with surgical procedures that often involve anesthesia and other drugs that affect the central nervous system. Lithium use should be stopped, with the approval of a healthcare provider, at least two weeks before a scheduled surgery.

Thyroid disease: Lithium might make thyroid problems worse."

Clearly Lithium isn't side-effect free, so it would be pretty reckless to add something like that to drinking water.

whatshisface4• 8 years ago

They see it as adding mind altering chemicals to water, which couldn't be further from truth.
We've added lithium to drinking water which has been linked to decrease in dementia.

Pick one.

jcranmer• 8 years ago

We've added lithium to drinking water which has been linked to decrease in dementia.

Who's the "we" here? The facility I worked at adds only the following chemicals in order: sulfuric acid (pH control, since the algal sex orgies in summertime causes pH to shoot way up), poly-aluminum chloride and a long-chained coagulant polymer (coagulants), ozone and sodium hypochlorite (disinfection), sodium hydroxide (more pH control), fluoride (the obvious), ammonium (for keeping the chloride around longer), and zinc orthophosphate (pipe corrosion control). There's also unused tanks for potassium permanganate (raw water control), alum (unused alternative for PACL as a coagulant) and calcium thiosulfate (ozone quench).

There's no lithium in the system at all--I've walked past every tank and pump.

ThePhysicist3• 8 years ago

Presenting the true positive rate without mentioning false positives seems problematic. I couldn’t find any information about that in the article, and they don’t link the paper / publication.

mtgx• 8 years ago

It's not problematic. It's on purpose and it's criminally misleading. If "AI companies" had to show their false positive rates, 99% of them would go bankrupt because then nobody would want to buy their bullshit.

samfriedman2• 8 years ago

The paper itself [0] claims 82% specificity and 100% sensitivity.

Meaning, 100% of positive cases in the test set were correctly identified, and 82% of negative cases in the test set were correctly identified. That corresponds to a 0% false negative rate and an 18% false positive rate. The test set is 40 imaging studies from 40 patients.

[0] https://pubs.rsna.org/doi/pdf/10.1148/radiol.2018180958

SolaceQuantum• 8 years ago

Wouldn’t 82% of negative cases being correctly identified mean that 18% of negative cases were incorrectly identified as positive(false positive), or in other words it will mistake 18% of healthy people as having Alzheimer’s, but will always identify an Alzheimer’s patient?

fencepost• 8 years ago

A false positive rate (currently) that high may preclude this being used as a widely available diagnostic test, but if used within a patient base already showing other signs of problems it might be more acceptable. After all, if you're going to see a memory specialist you're already thinking there's a problem.

pge• 8 years ago

This is a key point. If the false positive rate is 18% (as noted in other comments), and the rate of Alzheimer's is 36/100k[1], then for every patient the test correctly identifies, there are 500 patients that tested positive but did not have Alzheimers. There needs to be significant improvement before such a test becomes useful.

[1] that's mortality rate which I am using as a proxy for incidence rate, though the actual incidence rate is presumably higher (mortality rate from the CDC)

gooseus• 8 years ago

I'm not good at this (yet), but my understanding is that your prior for just this test would be the likelihood of having Alzheimer's based on the statistical occurrence within the general population.

After this test you now have updated priors which may not be good enough to make a certain diagnosis, but wouldn't it still be useful combined with another independent test that has similar rates, since you'd be using these updated priors?

bsder• 8 years ago

That depends. I suspect that this test would only be used on those with a family history of Alzheimer's or similar.

That alters the Bayesian priors quite a bit.

In addition, you can monitor the progress of someone who "tests positive" over time and see if that is really true.

Part of the issue with Alzheimer's treatments is that we may be intervening far too late. These kinds of tests may help that.

epmaybe3• 8 years ago

Firstly, I agree. You can't hide from that, and research should be upfront with the limitations of the work.

Other commenters have noted that the sensitivity is 100% with specificity 82% (false positive rate being 18%). I wanted to provide my own take on what that means for me as a future provider (I'm considering becoming a neurologist after medical school). I'm assuming to the data science crowd here, it's almost second nature to know what these numbers really mean in practice, but maybe I can provide some outside perspective.

When considering screening tests in medicine, we have to always balance the benefits and risks of conducting these tests on patients. We have to figure out how reliable a test is, and how patients benefit from earlier detection.

In this case, it's an imaging test, so noninvasive. However, it's a test that requires radiation (positron emission tomography or PET). So we already have to determine if that extra radiation dose is worth the increased risk for something like cancer. In an elderly population, it's almost certainly worth the scan given the radiation would not significantly impact their decades of accrued radiation damage.

Now, what happens if a patient has a positive test, 6 years earlier than they would normally? Let's assume for a moment that it was a true positive test. They could receive treatment earlier, which could improve their quality of life. They could adjust their risk factors to slow down the progression of disease. The treatment could have side effects, such as increased blood pressure. If we apply that to a large population scale, would we increase mortality associated with heart disease by treating patients with suspected Alzheimer's earlier?

Lastly, I wanted to touch on the effects of a screening test with a large amount of false positives. I would argue that a screening test should be allowed to have lots of false positives, simply because we can always do other tests (neuropsychiatric evaluation like MOCA score) to improve our accuracy of detection. However, if we do these other tests, we have to again balance the harms with the risks. Mammograms, for instance, could give false positives that could warrant a breast biopsy. In the case of Alzheimer's, a false positive could lead to stress and depression in an elderly population already at higher risk for stress related disease.

In short, screening tests are great if there's a way to do something about the disease. As far as I know, early prophylactic treatment (with current medications) of Alzheimer's does improve quality of life, slowing progression of the disease manifestations, with relatively few side effects.

citrablue• 8 years ago

You missed some downsides: anguish, more tests for something that isn't a problem, costs, possibly suicides. You may want to read up more on the downsides to breast cancer screenings for reference.

zaptheimpaler• 8 years ago

Woah is the radiation in an imaging test really so dangerous that you carefully consider whether its worth doing even once? Because uhh i get those pretty often

max762• 8 years ago

Quick algorithm that identifies all brain scans with Alzheimer's.

public bool HasAlzheimer(GlucosePetScan patientScan) { return true; }

woodandsteel• 8 years ago

Yes, but whatever the false positive rate is, my guess is it will decrease significantly in coming years. For instance, if a patient gets a positive rating, there might be another type of test that would eliminate most of the false positives. It seems to me quite unlikely researchers would not try to make improvements.

thegambit• 8 years ago

*Machine Learning

TaupeRanger2• 8 years ago

*Computational Statistics (a better name for both, given current methods, since it avoids anthropomorphism and sci-fi implications in uninformed readership).

crb002• 8 years ago

23–26 mSv of radiation. If the Alzheimer's doesn't kill you the PET scan will from a secondary cancer.

ceejayoz• 8 years ago

https://www.nrc.gov/reading-rm/doc-collections/fact-sheets/b...

The data show high doses of radiation may cause cancers. But there are no data to establish a firm link between cancer and doses below about 10,000 mrem (100 mSv – 100 times the NRC limit).

sameschker• 8 years ago

I found the publication that the article discusses.

https://pubs.rsna.org/doi/pdf/10.1148/radiol.2018180958

smartbit• 8 years ago

And the full publication can be found by entering https://doi.org/10.1148/radiol.2018180958 into sci-hub.

I often see https://outline.com/ links on HN, does HN allow sci-hub links?

gyvastis• 8 years ago

Thanks for the cool tool, didn't knew about it!

bougiefever• 8 years ago

My grandmother died from this, and it is a horrible death. I really fear becoming a victim to this disease. Anything that can help, even a little, is so encouraging.

zwieback• 8 years ago

How do they compare the AI performance to that of radiologists? The abstract says that radiologists aren't as good at spotting subtle, diffuse changes but what metric did they use to compare?

Eridrus• 8 years ago

You get historical PET scans from years ago and label them with whether they turned out to have Alzheimer's later in their life or not.

gesman• 8 years ago

"People never lie so much as after a hunt, during a war or before an election"

-- Otto von Bismarck

This quote need to be updated for modern times regarding performance of AI/ML models.

m15i• 8 years ago

I'm a co-author of this paper and happy to answer any questions.

Traster• 8 years ago

I find it really interesting that you used machine learning to do the diagnosis given the data you had available - scans seem to be naturally similar to images that we use neural networks on today so t he fit seems good. I was wondering though, did you have any ability to dissect the algorithm after it was trained and see what exact characteristics it was looking for?

The reason I ask is because there's some description of the processes we know indicate Alzheimers, but there could be new signs this algorithm has identified that could be applied more directly?

johnnyfived• 8 years ago

No questions here, but you guys might be interested to see the implementation I took with this, where I also used machine learning for Alzheimer's / dementia diagnosis.

https://github.com/jddunn/dementia-progression-analysis

I was using OASIS's public dataset so I only had ~150 images to work with, instead of ~2000. I used transfer learning from ImageNet's dataset to try and get usable results. I also had super limited testing (15-20 patients), but got ~60% accuracy with ~13% false positive rate.

It could be useful to apply those same transfer learning techniques in your team's model.

alexnewman• 8 years ago

This is a great thing, but i wonder if it's actionable. What are the insurance impacts?

mendelsd• 8 years ago

Interesting that glucose levels are used, since they are significantly affected by diet: https://biology.stackexchange.com/questions/55853/what-is-th...

AI Can Detect Alzheimer’s Disease Six Years Before a Diagnosis

Related Stories