Abstract
The clinical symptoms of schizophrenia are highly heterogeneous, with the most striking symptoms being cognitive deficits and perceptual disturbances. Cognitive deficits are typically linked to abnormalities in top-down mechanisms, whereas perceptual disturbances stem from dysfunctions in bottom-up processing. However, it remains unclear whether schizophrenia is primarily driven by top-down control mechanisms, bottom-up perceptual processes, or their interaction. We hypothesized that abnormal top-down and bottom-up interactions constitute the neural mechanisms of schizophrenia. Considering that autoencoders can identify hidden data features and support vector machines are capable of automatically locating the classification hyperplane, we developed an improved stacked autoencoder-support vector machine (ISAE-SVM) model for diagnosing schizophrenia based on resting-state functional magnetic resonance imaging data. A permutation test was used to identify the 213 most discriminative functional connections from the model's output features. Functional connections linking regions of higher cognitive functions and lower perceptual tasks were extracted to further examine their relevance to clinical symptoms. Finally, spectral dynamic causal modeling (sDCM) was used to analyze the dynamic causal interaction between brain regions corresponding to these functional connections. Our results showed that the ISAE-SVM model achieved an average classification accuracy of 82%. Notably, five resting-state functional connections spanning both cognitive and sensory brain areas were significantly correlated with Positive and Negative Syndrome Scale scores. Furthermore, sDCM analysis revealed weakened top-down regulation and enhanced bottom-up signaling in schizophrenia. These findings support our hypothesis that impaired top-down regulation and enhanced bottom-up signaling contribute to the neural mechanisms of schizophrenia.