Abstract
The heterogeneity of psychotic disorders leads to instability in subjectively defined diagnoses. This study used a machine learning framework termed common orthogonal basis extraction (COBE) to decompose electroencephalography-based functional connectivity (FC) in patients with psychotic bipolar disorder (PBD), schizophrenia (SCZ), and schizoaffective disorder (SAD) into individualized and shared subspaces. The results demonstrated that individualized FCs captured disease heterogeneity and predicted symptom severity more accurately than raw FCs, while shared FCs revealed diagnosis-specific abnormalities and achieved an accuracy of 79.30% in differentiating PBD, SCZ, and SAD. Furthermore, molecular decoding implicated regionally selective serotonin systems and astrocytes in the neurobiological differences among disorders, suggesting disorder-specific pharmacological targets. Critically, these findings were replicated in an independent cohort, confirming the effectiveness of the COBE framework in mining neurophysiological and molecular profiles of schizophrenia-bipolar disorder. These findings advance mechanistic understanding of psychotic disorders and offer a promising avenue toward objective, clinically relevant tools for psychotic evaluation.