The interaction network and potential clinical effectiveness of dimensional psychopathology phenotyping based on EMR: a Bayesian network approach.

Abstract

The current DSM-oriented diagnostic paradigm has introduced the issue of heterogeneity, as it fails to account for the identification of the neurological processes underlying mental illnesses, which affects the precision of treatment. The Research Domain Criteria (RDoC) framework serves as a recognized approach to addressing this heterogeneity, and several assessment and translation techniques have been proposed. Among these methods, transforming RDoC scores from electronic medical records (EMR) using Natural Language Processing (NLP) has emerged as a suitable technique, demonstrating clinical effectiveness. Numerous studies have sought to use RDoC to understand the Diagnostic and Statistical Manual of Mental Disorders (DSM) categories from a qualified perspective, but few studies have examined the distribution variations and interaction characteristics of RDoC within various DSM categories through retrospective analyses. Therefore, we employed unsupervised learning to translate five domains of eRDoC scores derived from electronic medical records (EMR) of patients diagnosed with Major Depressive Disorder (MDD), Schizophrenia (SCZ), and Bipolar Disorder (BD) at West China Hospital between 2008 and 2021. The distribution characteristics, interaction networks, and potential clinical effectiveness of RDoC domains were analyzed. Using non-parametric statistical tests, we found that MDD had the highest score in Negative Valence System (NVS) (4.1, p < 0.001), while BD exhibited the highest score in Positive Valence System (PVS) score (4.9, p < 0.001) and Arousal System (AS) (4.4, p < 0.001). SCZ demonstrated the highest scores in Cognitive Systems (CS) (5.8, p < 0.001) and Social Processes Systems (SPS) (4.6, p < 0.001). Through Bayesian network (BN) analysis, we identified relatively consistent interaction relationships among various RDoC domains (NVS → AS, NVS → CS, NVS → PVS, as well as CS → SPS; parameter range = 0.156 to 0.635, p < 0.001). Lastly, using logistic regression and Cox proportional hazards models, we demonstrated that AS was significantly associated with the length of hospital stay (-0.21, p < 0.05) and 30-day readmission risk (adjusted odds ratio [aOR] = 0.91, 95% confidence interval [CI] 0.91-0.99) to some extent. In conclusion, we suggest that the eRDoC characteristics varied in different DSM. By Bayesian Network, we found NVS and CS might be potential source in interacting with other system. Furthermore, CS, SPS and AS were associated with the length of stay and 30-days readmission, making them effective for predicting prognosis of psychiatric disorders.