T73. IDENTIFYING KEY VOXELS IN SCHIZOPHRENIA THAT ARE CORRELATED WITH AGE OF ONSET AND DURATION OF ILLNESS

Schizophrenia is a chronic, disabling mental disorder. Patients suffering from multiple degradations. Presently, several different brain regions are found to be involved in the neuropathology of schizophrenia, including limbic and temporal lobe, cingulate gyrus, and basal ganglia [1]. By applying the deep learning method in structural brain magnetic resonance images, an explainable deep neural network (EDNN) framework is used to identify the key structural deficits in schizophrenia [2]. We then sought to identify the correlation between demographic and cognitive profiles and structural deficits in schizophrenia.

We used the general linear model to examine predictors of clinical assessment scale in response to two different voxel integrity models for patients with schizophrenia.

The EDNN key voxels included 183 voxels which were trained by the structural MRI data, which is consisted of 200 schizophrenic patients and 200 age and gender-matched healthy control subjects. Brain MRI images were normalized and segmented into gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF) space. The clinical assessment data were obtained from the same group, including sex, age, onset age, duration of illness, digit span task and mini-mental status examination. Next, the average image intensity from identified key voxels was used as the response, and cognitive data as predictors to build a regression model. We also compared the model results with those obtained from anatomical parcellation with significant between-group differences in the image intensity.

In terms of its predictions to the integrity of grey matters using the linear regression model, the EDNN data yields 0.33 of R-squared value, and on the other hand, anatomical parcellation reaches 0.33 of R-squared value. We also found that the key voxels identified by the EDNN were significantly correlated to the age of onset and duration of illness.

Our results suggest that, at the statistical level, our EDNN dataset can derive comparable results using much fewer voxels. The structural deficit identified by EDNN model was mostly contributed by the age of onset and duration of illness, which is consistent with gray matter loss observed in the course of schizophrenia.