Bulimic-type eating disorders, like bulimia nervosa (BN) and binge eating disorder (BED), are characterized by a wide range of neurocognitive alterations. Neuroimaging studies have highlighted impairments in brain networks associated with self-control functions, including inhibition, reward processing, and homeostasis. However, case-control analyses that focus on a limited set of these domains frequently produce inconsistent results. Given the complex behavioral patterns characteristic for these disorders, particularly regarding eating behaviors, these findings suggest substantial variability and interaction at the neurobiological level, which likely underpins the heterogeneous nature of these disorders. This study explores the utility of a data-driven machine-learning approach to predict diagnosis and disease severity in subjects with either BN, BED or healthy controls at the individual level. We included 30 patients with BN and 30 patients with BED from a previously published dataset. To better capture complex network features and their interactions, and to extend traditional univariate analyses, we trained multiple predictive models using a combination of neuroimaging data (structural, resting-state, and task-based), behavioral performance measures, blood parameters, and psychometric questionnaires. Models are fitted with a 5-fold nested cross-validation procedure and integrated feature selection and hyperparameter tuning. We hypothesize that the multimodal integration of diverse feature types will improve discriminative accuracy when distinguishing between patients and controls, compared to models based on a single modality. Additionally, we expect that the contribution of individual features will differ depending on whether the model is predicting disease severity or diagnosis. The results of this study will be presented at the conference.