Reversing Alzheimer’s

(chris) #1

Hi everyone, Dale Bredesen, MD, PhD is on the cutting edge of reversing Alzheimer’s and other dementias.

Alzheimer’s is like cancer in that it’s not one single disease. That means there will never be a single-drug cure and clinical trials will always be of limited use.

One of the things I find most interesting about Bredesen is he’s using an algorithm to analyse the data he collects from patients. Is it a learning algorithm? I doubt it. I’ll know more early next month after I’ve attended his training course at the Buck Institute with my co-founding MD PhD.

If you’d like to know more, I’d highly recommend these three presentations:

Professor Dale Bredesen - Why Nutrition is the key to Alzheimer’s (part one) (part two) (part three)

If this is the kind of problem that gets you fired up, please get in touch. I will likely be hiring in the near future.


Deep learning with medical images
(Jeremy Howard (Admin)) #2

There’s also some terrific datasets available here: . They are neuro-imaging for Alzheimer’s, which is another important area for this affliction.


(jbrown81) #3

Thanks for sharing the talks Chris. Fascinating stuff.

On a related note: I’ve put together a dataset of structural MRI images compiled from the ADNI study Jeremy kindly pointed out. As it stands there are 3013 scans from 321 subjects (190K 96x96 2D images), who are about an equal balance of cognitively normal, mild cognitive impairment, and Alzheimer’s disease. The scans have been preprocessed and saved in numpy arrays and the metadata for each scan is in a .csv. I have a jupyter notebook where I’ve made a first pass at a convolutional classifier (CN/MCI/AD) and an autoencoder for unsupervised representation of images.

If anyone wants a crack at this data, or to put heads together on these brain image problems, I’m happy to share the data + notebook. Let me know!


(Jeremy Howard (Admin)) #4

It would be great to share it. I wonder if you could create a kaggle dataset and kaggle kernel ? Or is there some other good way to collaborate on this?

1 Like

(ben.bowles) #5

Pretty neat stuff! I wonder how you even start building a model trying to discriminate MRI scans. Enormous data! Can you use 3d convolutions? The other thing, I suspect you could a very strongly performing model just based on the size of the hippocampus. What would be really interesting would be if you could see, in individuals with mild cognitive impairment who do have any hippocampal atrophy (as of yet), the structure of the hippocampus (perhaps revealed in some way through 3d CNNs) was markedly different in mild cognitive impairment, even if it was overall the same size. That might actually be science worthy :wink:

1 Like

(chris) #6

Yes! @jbrown81 Kaggle it!

How much does an MRI cost in the US? What about elsewhere in the world? Could you predict the MRI using a less expensive test?


(jbrown81) #7

I’m on my way to getting a Kaggle dataset set up for the ADNI images. Data + labels should be ready for download early next week, I’ll send out the link as soon as it’s up.


(jbrown81) #8

Finally getting around to replying in the wake of the Thanksgiving. 3d convolutions are definitely a possibility; agreed that they could learn all sorts of interesting things about brain region shape (eg in hippocampus) being a risk factor before volume loss is apparent. There are a couple reasons why training on 2d images might be worth investigating before going to 3d:

  1. a lot of clinical MRI scans are only on intermittent 2D slices, since it’s cheaper/faster than a full brain scan. it could be quite valuable to have a Alzheimer’s classifier that can do a good job from as few 2D slices as possible

  2. if you train a model on 2D images instead of 3D, you increase your dataset size by ~60x, since you’re classifying 2D images instead of whole brains. of course, the tradeoff is that you are missing out on information when you go from 3D->2D, so some sort of aggregate measure across 2D images may be useful.


(jbrown81) #9

How much does an MRI cost in the US?
prices vary a lot, but the rough range is $500 - $4000, of which you don’t pay more than $50-$100 if you’re insured.

What about elsewhere in the world?
cheaper. maybe $400 in france or $100 in india, according to a couple quick google searches. again, estimates vary but they’re basically overpriced in the US. the reasons are more a function of insurance and politics than technology.

Could you predict the MRI using a less expensive test?
good question! maybe with mobile health sensors, analysis of speech patterns, sleep patterns, etc - i just don’t think we know yet. we’ll probably need MRI and PET scans for a while to validate a lot of the cheaper but potentially noisier signals.

1 Like

(Jeremy Howard (Admin)) #10

I suspect a 3d model would work well even in this case.

1 Like

(jbrown81) #11

The ADNI data is now up on Kaggle at


(jbrown81) #12

One more update on the ADNI data, here’s a jupyter notebook for working with the data (loading images + labels, using keras to set up and train a convnet, viewing images, diagnosing good + bad classifications)


(Jeremy Howard (Admin)) #13

This is great! Have you considered writing a blog post about it? I think it would be extremely popular and bring a lot more people to look at this problem.

Also - have you tried creating the notebook directly on the kaggle ‘kernels’ platform? If you do, it’s a bit easier for others to use it - and you don’t have to worry about loading in the data.

1 Like

(jbrown81) #14

Thanks! I will definitely blog it (and kernel it) in the next couple days - have a couple cool analyses of the data I’m half way through writing up.


(Jeremy Howard (Admin)) #15

A humble request: we’d love to highlight some projects related to this class - so if you happen to have a chance to mention that you’re part of the course, we’d be much obliged (we’re hoping to use this to encourage people to take the MOOC - amongst other things so that there’s plenty of folks qualified for part 2 next year!)

(Absolutely no obligation, of course)


(mattobrien415) #16

This is all great stuff. I work at the Memory and Aging Center here at UCSF where we study Alzheimers and FTD.

I’m trying to continue work on a project that has been discontinued a while back; a researcher did some good work predicting alzheimers from conversation (looking at both the characteristics of both the audio signal and the semantic features of the actual language).

It looks like ADABoost did the best, but they stopped short before doing any deep learning. I’m trying to get the data and code but the guy left and we’re having trouble getting it! :unamused:

If this all does come together, I’d love to write up a blog post or something and point folks towards the course.

(We’ve also got tons of MRI data that would be nice to check out at some point)


(Jeremy Howard (Admin)) #17

That would be amazing @mattobrien415 . If you manage to get that dataset together, I hope you’ll let us know here on the forums so we can all help you make the most of it! :slight_smile:


(chris) #18

I’ve just got back from three days of training at the Buck Institute for Research on Aging with one of my doctors who has a PhD in neuroprotection. In case anyone is interested, I’ll summarise what we learned:

  • Dale Bredesen is a researcher and a clinician. He has published both case studies and on the basic science of AD.
  • Mild cognitive impairment is reversible but once you’re in the nursing home, you’re probably not coming out.
  • AD is like fever in that it’s not acceptable to diagnose without stating a cause.
  • Bredesen’s protocol is increasing hippocampal volume, something previously thought to be impossible.
  • The protocol is primarily diet and lifestyle based, but also calls for a battery of tests to look for chronic infections, nutrient deficiencies, heavy metal and biotoxin exposure.
  • Brain volumetrics are being tracked using MRI and the Neuroreader software. The software transforms an image that the clinicians don’t understand into a large table of volumetrics that the clinicians also don’t understand, but it doesn’t matter because the MRI doesn’t change the protocol anyway. Radiologists get super cranky if you talk to them about deep convolutional neural networks :slight_smile:
  • Dale created if-then-else software that reduces the dimensionality of hundreds of biomarkers down two five subtypes of AD but there’s tremendous overlap between the types and even with other dementias.

Our goal is to help as many people as possible and at the moment I’m not sure how how MRI achieves that because it doesn’t change the treatment. Would love to hear what other people think about this!


(David Gutman) #19

Don’t want to argue, but I’d be very skeptical of this (and FYI deep convolutional networks don’t make me cranky).

The protocol is primarily diet and lifestyle based, but also calls for a battery of tests to look for chronic infections, nutrient deficiencies, heavy metal and biotoxin exposure.

Pseudoscience does make me cranky.

So does seeing young people with horrible strokes after trips to the chiropractor, horrible metastatic cancer that would have been curable had the patient seen an “allopathic” rather than a “homeopathic” doctor, and measles outbreaks due to people refusing to vaccinate.

The issue with Alzheimer’s is that the only real way to diagnose it is autopsy. There are some PET tracers that show promise for catching the disease when it still might be reversible (e.g. Amyvid), but those are still very new.

Admittedly I used a biased google search to find this, but…


(chris) #20

Hi @davecg, if pseudoscience makes you cranky, then I’d suggest you stick to PubMed when retrieving science. For your convenience, here’s a list of Dr Bredesen’s publications:[Author]

Dr Bredesen is a traditionally trained medical doctor with published research dating back decades.

The Journal of Neuroscience is an important journal; is not.