Enhancing Prediction of Individualized Antipsychotic Outcome with fMRI-EEG Feature Integration in First-Episode Schizophrenia: A Real-World Study

Predicting treatment efficacy in schizophrenia within real-world settings is of great clinical significance but challenging. fMRI and EEG reveal distinct spatial and temporal characteristics of brain information processing. Therefore, developing predictive models that integrate spatiotemporal multimodal features and accommodate the complexities of real-world environments is crucial for achieving accurate treatment outcome predictions and personalized therapy. Ninety first-episode, drug-naive schizophrenia patients underwent fMRI and EEG at baseline and received 6–8 weeks of single antipsychotic treatment in a naturalistic setting. Clinical symptoms were evaluated using the Positive and Negative Syndrome Scale (PANSS) at baseline and post-treatment. Entropy metrics reflecting information processing capacity were calculated from BOLD signals as fMRI features, while key ERP components (P50, N100, P200, N200, and P300) representing different cognitive stages were extracted with their amplitudes and latencies as EEG features. LASSO regression model was used to assess the predictive power of unimodal and multimodal features for PANSS score reduction. The multimodal model outperformed unimodal models in predicting improvements in PANSS total score (R = 0.440, P < 0.001), negative symptom (R = 0.328, P = 0.002), and general psychopathology (R = 0.449, P < 0.001). Key features in the multimodal model included WPE from the medial frontal gyrus, supplementary motor area, amygdala, caudate, pallidum, and thalamus, along with ERP characteristics like P50 ratio, N100, N200, P3a latencies, and P200 amplitude. These features were not significantly correlated, highlighting their complementary roles in information processing as key to the multimodal model's improved performance. This study demonstrates that multimodal fusion prediction model effectively integrates brain information processing features across different dimensions, significantly improving individualized prediction of treatment outcomes. The results hold important value for its translational application in precision psychiatry within real-world settings.