IMA-MoE: An Interpretable Multimodal Mixture-of-Experts Framework Reveals Neurobiological Features of Binge Eating Disorder

IMA-MoE is an interpretable multimodal perception mixture-of-experts architecture. By integrating multi-dimensional data such as neuroimaging, behavioral, hormonal, and demographic data, it enables flexible modeling of cross-modal dependencies. This framework performs excellently in distinguishing patients with binge eating disorder from healthy controls on the ABCD dataset and reveals gender-specific prediction patterns, providing a new perspective for neurobiological research on binge eating disorder.

Diagnosis Dilemma of Binge Eating Disorder

Binge eating disorder is the most prevalent eating disorder globally, but current diagnosis relies on symptom descriptions rather than biological mechanisms, with limitations such as difficulty in early detection, limited intervention methods, and neglect of individual differences.

Limitations of Existing Research

Traditional research faces issues such as hypothesis-driven parametric models being difficult to capture complexity, unimodal analysis ignoring cross-modal effects, and small sample homogeneity limiting generalization ability.

Core Architecture

1. Tokenized Multimodal Representation: Encodes each type of measurement (neuroimaging, behavioral, hormonal, demographic) into independent tokens, preserving modal characteristics and supporting flexible modeling.
2. Mixture-of-Experts Mechanism: Achieves specialized processing, sparse activation, and capture of complex non-linear relationships by activating specific expert sub-networks.
3. Modal-Aware Routing: Routing decisions consider modal sources, identify key modal combinations, and adapt to modal missingness.

Token Importance Mechanism

Provides feature-level attribution, individualized interpretation, and interpretable analysis aligned with clinical practice.

Classification Performance

In the task of distinguishing patients with binge eating disorder from healthy controls, IMA-MoE significantly outperforms baseline methods, showing higher accuracy, robustness, and generalization ability.

Interpretability Findings

1. Gender-Specific Prediction Patterns: Hormonal measurements contribute prominently in female samples, while neuroimaging features are more important in male samples.
2. Multimodal Synergistic Effects: The combination of neuroimaging and behavioral indicators shows predictive value, and demographic variables regulate the importance of other modalities.

1. Precision Diagnosis: It is expected to develop biomarker-based screening tools to enable early detection.
2. Gender-Specific Intervention: Female patients can benefit from hormone regulation interventions, while male patients need interventions targeting neurocognitive functions.
3. Interpretable AI Applications: Enhances understanding of disease mechanisms, guides the discovery of therapeutic targets, and builds clinicians' trust in AI.

The current study has limitations such as lack of longitudinal analysis, insufficient causal inference, and need for verification in clinical translation. Future work needs to conduct longitudinal follow-up studies, combine experiments to verify causal relationships, and promote clinical translation efforts.

IMA-MoE represents an important advancement in the field of mental health AI. It not only improves classification performance but also reveals clinically valuable insights such as gender-specific prediction patterns, providing new ideas for precision psychiatry. With the evolution of multimodal technology and AI methods, this framework is expected to play a role in research on a wider range of neuropsychiatric disorders.