Autism and HIV: where mathematics can be misleading

An article in the Journal of Neuroscience has received wide publicity because it refers to a method of diagnosing autism in adults using brain scans, with 90 percent accuracy. He uses the methods of pattern recognition, training a computer to distinguish between tests for people with and without autism. 

 The resulting test is much cheaper than usual, as all this sounds promising. But the 90 percent accuracy a focus on the question "What is the frequency of this right to test" The fundamental question is: "How often is it wrong button?
Why? Clients.
Approximately one in 100 suffer from autism. Suppose we apply the test of a million people, of whom 10,000 were in fact autistic. Then, among those with autism, we would have 9000 and 1000 diagnosis right is wrong - not bad for a cheap test of 15 minutes. But among the 990,000 without conditions, not the test produces good diagnosis, and falsely diagnose autism in 99,000 cases. So 99.000 making a total of 108,000 people with autism, do not suffer from the condition. It is a mark of about 9 percent.
This is a rough and ready calculation. With regard to the details of the results changed somewhat - about 5 percent.
If there are independent reasons to suspect that someone autism, and false negatives - the failure to state even if it is available to diagnose - is the major concern. But if you're a part of the population, most of which are not autistic test, then it is false positive and causing problems. The same questions apply to a relatively rare disease. In the early days of AIDS, a couple committed suicide a very accurate test shows that HIV-positive, while in reality, the test resulted in a large proportion of false positives.
There are other problems with the study of autism. The computer is trained to distinguish the brain scans of people with and without autism consultant least five different aspects of the analysis and development of the mathematical combination of them were most closely with the presence or absence of autism. But the features are all major scale - the thickness of the cortex, the complicated way the brain surface, and so on. These features are specially adapted not a specific condition to diagnose, it's a bit like trying to diagnose chickenpox by looking weight, height and hair color.
In addition, the number of people involved was small: 20 with autism, 20 without. This small group, it is unclear whether an association that arises is important. You can use a computer to learn by using pictures of the cat family, and it will be calculated according to the combination of size, color, length and whisker better detect autism in its owner. There are so many possible combinations in all likelihood one of them looks very good play. But try it on another group of people, and chances are it will fail.