Neural network models of schizophrenia

Abstract

There is considerable neurobiological evidence suggesting that schizophrenia is associated with reduced corticocortical connectivity. The authors describe two neural network computer simulations that explore functional consequences of these abnormalities. The first utilized an "attractor" neural network capable of content-addressable memory. Application of a pruning rule that eliminated weaker connections over longer distances produced functional fragmentation and the emergence of localized, "parasitic" attractors that intruded into network dynamics. These pathologies generally were expressed only when input information was ambiguous and provide models for delusions and cognitive disorganization. A second neural network simulation examined effects of corticocortical pruning in a speech perception network. Excessive pruning caused the network to produce percepts spontaneously, that is, in the absence of inputs, thereby simulating hallucinations. The "hallucinating" network also demonstrated subtle impairments in narrative speech perception. A parallel study of human patients found similar impairments when comparing hallucinating patients with nonhallucinating patients. In addition, the authors have used transcranial magnetic stimulation (TMS) to directly probe speech perception neurocircuitry in patients with these hallucinations. As predicted by the neural network model, the authors confirmed that "suppressive" low-frequency TMS reduces auditory hallucinations. Neural network simulations provide empirically testable concepts linking phenomenological, cognitive, and neurobiological findings in schizophrenia.