IBM's AI Can Predict Schizophrenia With 74 Percent Accuracy By Looking at the Brain's Blood Flow (engadget.com)

Andrew Tarantola reports via Engadget: Schizophrenia is not a particularly common mental health disorder in America, affecting just 1.2 percent of the population or around 3.2 million people, but its effects can be debilitating. However, pioneering research conducted by IBM and the University of Alberta could soon help doctors diagnose the onset of the disease and the severity of its symptoms using a simple MRI scan and a neural network built to look at blood flow within the brain. The research team first trained its neural network on a 95-member dataset of anonymized fMRI images from the Function Biomedical Informatics Research Network which included scans of both patients with schizophrenia and a healthy control group. These images illustrated the flow of blood through various parts of the brain as the patients completed a simple audio-based exercise. From this data, the neural network cobbled together a predictive model of the likelihood that a patient suffered from schizophrenia based on the blood flow. It was able to accurately discern between the control group and those with schizophrenia 74 percent of the time. What's more, the model managed to also predict the severity of symptoms once they set in. The study has been published in the journal Nature.

Comments

Fire analogy

[John Allsup]

Consider fire. Consider fire fighting, fire detection, and fire prevention.

There are many well known ways of using either heat, or presence of smoke in the air, to indicate a high likelihood that there is a fire in a region of a building. But these detection methods do not tell you anything about how the fire started. Combining information from many detectors across a large building can tell you about how a fire is spreading, but not about how a raging fire _might_ spread.

That people who have had some episodes labelled as 'schizophrenia' leaves common tell-tale signs detectable in this way is good to know, from a research point of view. But just like the 'fire analogies' above, where multiple similar looking fires, with similar results, can start in markedly different ways, the similar features in brains of people diagnosed with 'schizophrenia' only tell _part_ of the picture. As for how their brain came to be that way, such evidence can be likened to evidence that a fire in one room is a common cause of a fire in a neighbouring room. Things like 'chemical imbalances' are often touted as the 'cause' of things like bipolar disorder, or schizophrenia, rather than as a link in a causative chain. (This to me has always seemed as silly as saying that the pictures on your TV are _caused_ by electrical fluctuations in the aerial.)

In general, I think people working with mind and brain tend to overgeneralise, exaggerate, and oversell the consequences of their observations. This is further compounded when potential counter-evidence is 'defended against' and 'argued away', as happens between different factions of the mental health profession. People want things to be straightforward and simple, as if treating cuts and broken bones, and often inadvertently assume things are that simple before proceeding with studies whose results rely upon statistical reasoning which is contingent upon various assumptions uniformly holding across the population being studied... and then generally don't make clear their assumptions. People then read peer-reviewed research, and assuming a far simpler and more uniform picture than the evidence warrants.

Meaningless Summary Statistics

[goombah99]

Saying something is 74% accurate without stating false positive and false negative rates falls apart for rare diseases.

Here's an example: I have actually have a better method that can distingish between a control group and the real cases with 98.8% accuracy. I'm not kidding. All I do is I always say the person does not have the disease. Since 98.8% of people do not have it, I'm automatically correct 98.8% of the time.