Multi-task deep learning model to disentangle shared and unique brain functional changes associated with illness severity and cognitive functioning in schizophrenia

Individuals with schizophrenia experience significant cognitive impairments and alterations in brain function. However, the shared and unique functional regions underlying cognition deficits and symptom severity in schizophrenia remain poorly understood. We design a novel interpretable graph-based multi-task deep learning framework to enhance the prediction accuracy of schizophrenia illness severity and cognitive functioning measurements simultaneously by using functional connectivity, and identify both shared and unique brain regions associated with these clinical parameters on 172 subjects from COBRE and a local hospital. Our framework outperforms both single-task and state-of-the-art multi-task learning methods in predicting four PANSS subscales and four cognitive domain scores. Shared regions implicated in both illness severity and cognitive deficits include the supplementary motor area, middle temporal gyrus, and primary and secondary visual cortices. The Wernicke’s and Broca’s areas, and posterior cingulate are more strongly associated with symptom severity, while the superior and inferior temporal gyri, and superior parietal cortices are primarily related to variability in cognitive functioning. These findings are replicable across datasets and confirmed by meta-analysis at both regional and modular levels. Our study provides insights into the neural correlates of illness severity and cognitive implications, offering potential targets for further evaluations of treatment effects and longitudinal follow-up.