Classifying schizophrenia using functional MRI and investigating underlying functional phenomena.

Background: Existing studies have revealed functional abnormalities in certain brain regions of patients with schizophrenia (SZ), but the relationships between these abnormalities and their impact on disease progression remain unclear.

Methods: Fifty-six patients with SZ and 56 healthy controls were included. Based on resting-state functional magnetic resonance imaging, we analyzed fractional amplitude of low-frequency fluctuations (fALFF), regional homogeneity (ReHo), and degree centrality (DC). Statistically significant metrics were selected as features, and machine learning models were used to distinguish between patients and controls. Analyze the importance of features in the optimal model. The Louvain community detection algorithm and structural equation modeling were used to investigate community relationships and potential causal effects.

Results: The average prediction accuracy of various ML classifiers reached 0.9241 by fALFF, ReHo, and DC values. The SVM model have the highest performance with an accuracy of 0.9464. Abnormal ReHo in the right middle frontal gyrus contributed most to this optimal classifier and participated in the direct impact on SZ. All the features we analyzed ultimately constituted two functional clusters (FClus), which exhibit internal causal influences. FClus1 had a positive influence on SZ, with the cascade starting from abnormal fALFF in the right inferior temporal gyrus. FClus2 had a negative influence on SZ, with the cascade starting from abnormal fALFF in the left fusiform gyrus.Abnormal fALFF in the right caudate nucleus, degree centrality in the right angular gyrus, and ReHo in the right lentiform nucleus do not have a causal impact on the disease.

Conclusions: We identified interactions among features within FClus that potentially influence the onset and progression of schizophrenia, including epicenter phenomenon of FClus, FClus for inhibiting schizophrenia, and abnormal function of brain regions without direct impact. Additionally, we believe that the contribution of features to the disease classification model may indicate the size of their direct impact on the disease, not necessarily their importance in the disease process.