Diagnosing schizophrenia

An artificial brain aims to pick up early signs of schizophrenia

A controversial computer test for schizophrenia has been astonishingly accurate in early trials. Its inventors say the test, which is based on a learning system called a neural network, could even diagnose the condition before subjects show any symptoms.

Early trials have shown the program to be 100 per cent accurate. If further trials are as successful, it could allow early treatment of the disease, before it has had time to progress. The system was developed at the University of British Columbia in Vancouver by Peter Liddle, a psychiatrist specialising in brain imaging. It uses a neural network to analyse the brain scans of patients, looking for telltale characteristics in cerebral blood flow. "One of the big challenges with schizophrenia is the diagnosis. It can take several years for it to be made clear," says Liddle. "Being able to make a reliable diagnosis early can help to optimise the outcome." But some researchers are sceptical that brain imaging can provide all the answers. "Nobody has ever found a specific brain abnormality that all schizophrenics have that nobody else has," warns Robin Murray of the Institute of Psychiatry in London.

Liddle's system relies upon recent evidence suggesting that certain parts of the brain are disrupted in people with schizophrenia. In addition, activity at these sites differs between patients and healthy controls while they are carrying out verbal memory tasks.

According to Liddle, these anomalies might reflect the underlying causes of schizophrenia, making it possible to detect the condition much earlier. But the differences tend to be subtle, he explains. "This method seems to have the ability to pull out relatively complex patterns that the naked eye wouldn't be able to see," he says.

Ironically, the neural network is a computer program modelled on the human brain. It learns from experience, just as people do. Clusters of software processors, called nodes, are designed to behave like brain cells, weighting different attributes from an array of inputs according to their importance. The network has to be trained on known sample inputs, and the weightings are adjusted to produce the desired result. The program then responds to novel inputs in the same way. Liddle's neural network examines blood flow in the temporal and parietal lobes, looking for what its training has told it are the idiosyncrasies of schizophrenia. His network was trained on positron emission tomography scans taken from seven schizophrenia patients and two healthy subjects. After looking at scans from four healthy subjects and nine patients diagnosed as having schizophrenia, it was able to differentiate between them with 100 per cent accuracy. "This has a really significant potential," says Pat Levitt, a neuroscientist at the University of Pittsburgh in Pennsylvania. He says there are different types of schizophrenia-which are treated in different ways-but it's difficult to categorise patients. If doctors could use the program to classify patients, they would be able to treat them with a greater degree of certainty, he adds.

Author: Duncan Graham-Rowe

More at: Biological Cybernetics (vol 84, p 117)

New Scientist issue: 24th February 2001

Please mention New Scientist as the source of this story and, if publishing online, please carry a hyperlink to: http://www.newscientist.com