Schizophrenia is a psychiatric disorder best known by its intense behavioral symptoms. Novels and television shows have well characterized the more obvious issues like auditory hallucinations, paranoid delusions, and disjointed patterns of thinking. To date, however, neuroscientists have been unable to fit this diverse group of symptoms—as well as others like lack of affect, disorganized speech, and cognitive decline—into a single, cohesive theory that can explain both the cause and the development of the disorder. Researchers from the University of Texas at Austin and Yale University hope that the use of a neural network nicknamed DISCERN may offer better insight into how problems in learning and excess dopamine release can confuse the way schizophrenics remember language and events, eventually building up to psychosis.
Excess dopamine and the hyper-learning hypothesis
Researchers have known for decades that people with schizophrenia produce too much of the neurotransmitter dopamine. Most first-line drug treatments for the disorder block dopamine receptors to control hallucinations and other psychotic symptoms.
“There’s a lot of evidence that some kind of dopamine imbalance is important in schizophrenia,” says Uli Grasemann, a graduate student at the University of Texas at Austin. “The current theories suggest that too much dopamine may result in the brain giving too much significance, or salience, to different experiences. Then the brain learns from those abnormally encoded experiences, which then somehow gives you the symptoms.”
Grasemann, his advisor, Risto Miikkulainen, and Yale University School of Medicine psychiatrist Ralph Hoffman hypothesized that this excess dopamine and salience leads to “hyper-learning,” or an inability to discount and forget irrelevant information. Given the enormity of data the brain encounters daily, this hyper-learning, and the resulting abnormal connections and memories made from it, might eventually overwhelm a person to the point of psychosis.
The DISCERN model of schizophrenia
A neural network is a computer model, or program, designed to mimic aspects/behaviors of brain cells and their signaling patterns. The group created a network that can learn natural language. The model, named DISCERN, was then taught two different types of stories: everyday type of events and more outrageous tales that you might see in a novel or movie.
“The personal stories were told with an 'I,' to be like one’s personal experience. They were simple things like meeting a friend for a drink or getting a speeding ticket,” says Grasemann. “The other set of stories were straight out of the movies. They were all about gangsters and terrorists.”
Both types of stories were fed into the network and assimilated much the way the brain encodes verbal information by recording statistical relationships between the words, sentences and plots. Over time, the system learns the stories and can correctly reproduce them as output. But the researchers found that when they tweaked the model to hyper-learn, or attach more salience than necessary to information in the stories, the model’s output changed dramatically, making associations and connections between the agents and events in unique and disturbing ways.
“The stories were grammatically correct. It’s the same in schizophrenics—a schizophrenic’s grammar usually isn’t all that worse than a normal person’s grammar. But the context completely changed. They mixed up the different stories in ways that didn’t make any sense in a global context,” says Grasemann. “The model even started inserting itself into the more movie-like stories, even claiming to be responsible for a terrorist-bombing and things like that.”
That insertion of self into the different stories was of particular interest to Hoffman. He believes that understanding "agent sliding errors," or how the self or people closest to you can be erroneously associated with different unrelated events, may help scientists better understand how paranoid delusions form.
“You can imagine if a human being had all these confusions about their boyfriend or father-in-law how they might eventually conclude that there is some kind of plot,” Hoffman says. “You can see how, over time, they might read in different meanings to what these people say and do.”
He suggests that DISCERN’s hyper-learning simulates the early stages of schizophrenia. What’s more, the model suggests one doesn’t have to be in this hyper-learning state very long to produce psychosis.
“I would say that you only need to be in a state of hyper-learning for a week, maybe 5 days, to produce dramatic and devastating effects in terms of memory organization,” he says. “So if we could get to the point where we were able to detect hyper-learning early and potentially interrupt it in human patients, it really could be life-saving for them.”
Schizophrenia, beyond dopamine
Jeffrey A. Lieberman, the chairman of the department of psychiatry at Columbia University’s College of Physicians (and Dana Alliance for Brain Initiatives member), says that while this study is imaginative and well-orchestrated, it’s still very preliminary.
“From a theoretical standpoint, neural networks are very well suited to study brain function and dysfunction,” he says. “But the human brain is, by orders of magnitude, so much more complex than any other system in the body. It’s going to be hard to account for all the different variables, especially when you’re talking about a disorder like schizophrenia.”
While Hoffman concedes the point and says he believes there is more at work in schizophrenia than just excess dopamine, he believes this model opens up a “whole wealth of possibilities” for future study.
“For me, the model locates the illness at a much higher level than we’ve thought about it before, particularly pertaining to episodic or biographical memory. To me it’s something that feels intuitively correct,” he says. “It suggests schizophrenia is not just a peripheral problem but something that affects the way experiences are registered, stored, reproduced and interpreted. And that’s very exciting.”