Ibm and university of alberta use ai to predict schizophrenia with 74 percent accuracy

http://betakit.com/ibm-and-university-of-alberta-use-ai-to-predict-schizophrenia-with-74-percent-accuracy/

5 Likes

That’s good, with a database of less than 100 people. Imagine if there were 1,000 or 10,000.

I hope this will not stop there and they will be able to not only detect it, but tell which medication will be most effective. It really seems like the technology is there, only needs the data and the right algorithm. Isn’t this data sharing what things like the European brain project and the US BRAIN Initiative is for?

2 Likes

wow look @everhopeful

1 Like

You just beat me to it. It came in my rss feed.

I’m in shock.

1 Like

This is very promising! But remember that the base rate (prevalence) of schizophrenia is very low, so it’s probably not very useful at the moment. If 74 % of the positive predictions are correct, then 26 % are not, and that means for every 100 people tested, assuming a prevalence of sz of about 1 %, 26 will be wrongly predicted to develop schizophrenia and 0.74 will be correctly predicted to develop it. I’m assuming this is what they meant by 74 % accuracy. I will see if I can find the abstract of the study.

Edit: I found the research paper, but didn’t read it very closely. This is not my field and I was not able to comprehend all of it. But it does seem like this is roughly what they mean (they said it was 74 % effective at distinguishing people with schizophrenia from healthy controls). So it’s promising, but not quite good enough yet to be very useful. Hopefully, it will spark further research into the neural correlates of schizophrenia and computer modelling to predict and understand the illness.

1 Like

What’s so amazing is it could get to 74% accuracy with a sample size so small. Less than 100 people! Increasing the sample size should really help with accuracy.

1 Like

We’d probably need new knowledge first to see the kind of improvements that would make this useful for predicting who develops schizophrenia, though. With a base rate this low, a test would need very high accuracy to avoid high numbers of false positives.

What sort of new knowledge?

Thinking of how they teach any AI to recognize things, whether it’s a self driving car to recognize a pedestrian, or an image search to recognize a chair; they give it hundreds of thousands of images, if not more.

How powerful could this be, combined with blood and genetic data? Really powerful.

Here’s an article on a blood test for sz. A computer is certainly capable of cross referencing that witha scan and becoming highly accurate.

1 Like

New knowledge about such things as brain mechanisms, computer modelling and AI learning. We’re still quite a way from having AI that approaches the capabilities of the human brain when it comes to learning, and our knowledge of how the brain functions is also still very young. Also, with the amount of data that can be used in AI learning and computer models, more powerful computers could be a huge boon. Quantum computers could revolutionize AI. I don’t think more subjects alone would be enough for the drastic improvements this technology needs to be useful for predicting the development of schizophrenia. A 74 % accuracy over a random hit rate of 56 % (if I remember correctly from the article), is far from the accuracy we need for this application of the technology. We would need a false positive rate that is much, much lower. A 99 % accuracy would still be too low.

That depends on how much predictive power would be gained by the addition of new tests. Until we understand the causal mechanisms better, just adding new tests with moderate to low correlations with the development of the disease is highly unlikely to drastically improve the predictive power of the test battery. The gain from each new test added is likely to be diminished by intercorrelation, and because of the base rate, the total predictive power needed would be huge compared to predictive power of current tests.

This is just for diagnosis though, although adding data about meds response could make it a treatment tool.

We don’t have to know why a specific structure of the brain leads to sz for an AI tool to be able to see “when the blood flow/structure of the brain looks like this, that’s schizophrenia”

We don’t have to know how someone broke their arm, or how exactly bones heal to look at an xray and say it’s broken. It’s just sz is much more visually subtle and that’s where a computer has the edge to compare thousands of scans.

I’m truly amazed they could do it with what, 47 sz scans?

Yes, that was what I was talking about now. It could have many potential uses besides this. It could maybe be used as a tool for learning more about the illness, or tweaked, as you suggest, to become a treatment tool.

No, I realize that. That’s sort of why this research was done in the first place. But we shouldn’t expect this to be a magical solution to the problems of diagnosing and treating schizophrenia. The technology is just not there yet, and that’s not just because of the sample size. AI learning is not that impressive yet, the amount of data that could be analyzed in different ways is way more than even our most powerful computers can handle at this point, and so far, we do not have many reliable ways to predict whether someone will develop schizophrenia or not. I don’t think this will revolutionize psychiatry just yet. But hopefully some time in the future, when we find better predictors of schizophrenia, have better models and computers that can handle vast amounts of data, and/or have developed more impressive AI.

1 Like

Yeah I don’t think this will revolutionize sz diagnosis tomorrow. As fast as AI and computing power is increasing, I doubt it will take too long though. I’ll guess that human resistance, or just lack of funding and cooperation will be a bigger factor in delaying its implementation than further development/refinement of the technology.

This type of tool already has some application in bipolar, but it will take many years for that to become standard of care.

2 Likes

You never hear much about Kelly Clarkson anymore.