[ScottBurson]

It's great to see people working on this problem, and I wish you the greatest success.

I'm curious what you're comparing your results against to determine their accuracy. I thought there was no way to get a confirmed diagnosis of Alzheimer's short of an autopsy. (Obviously I'm no expert -- just going by what I've read.)

[Padideh]

You are absolutely right. The definitive test to confirm Alzheimer's pathology in the brain is by autopsy, however there are pretty extensive clinical tests available that get very close to the diagnosis. The accuracy of our software is determined by testing a population of people with clinical Alzheimer's disease, against their age-matched normal controls.

[sova]

You can also have a conversation with the afflicted human being and find out. Some days are more lucid than others, but on the whole there can be a 70-90% "fog" of dementia/alzheimer's that is apparent even without a brainslice.

[hprotagonist]

In general, you can't discriminate between dementia and alzheimers with a patient interview.

[Padideh]

Correct. Alzheimer's disease is the most common cause of dementia (~80%). There are other dementia's such as FTD (frontotemporal dementia), DLB (dementia with Lewy body), vascular dementia, etc.

[manmal]

I think they are using those 3000 brain scans of people who progressed to Alzheimer's - questions arise: what part of that was the training data, what part was the test set, and did the sets overlap?

[Padideh]

This is not a machine learning algorithm, so no training is required. The software processes every MRI scan (a full set acquired in one session) independently of other scans. The algorithm quantifies microscopic structural distortion at every voxel (3D pixel) of the MRI. Statistical models can be used on a population, but only to test the performance of the algorithm, not to train it.

[pjschlic]

I get you might not have officially trained a model, but similar model evaluation questions might apply to your diagnosis system.

For example, did you do any hand tuning of thresholds? Did you do this tuning while looking at the scans, or did you design it without looking at ANY of the mentioned 3000 scans - and are you are basing your 90% accuracy on diagnosing a sample which is how representative of the population at large?

Is there a further description of this 90% statistic - do you have precision / recall and some sort of description of possible biases in this 3000 sample set? Based on another answer, it sounds like half are alzheimer's and half are 'similar' but non-alzheimer's. I don't want to be flippant, but I am always sceptical of 'accuracy' in medical diagnostics where a simple diagnosis of 'false' is 95% accurate.

Edit: I see you answered this second section in another comment, nevermind.