An expert routing-driven graph framework from tongue multi-task features to liver fibrosis prediction

Abstract

Liver fibrosis is a global health concern where early detection is crucial for timely intervention. Traditional Chinese Medicine tongue diagnosis offers a promising non-invasive assessment, but suffers from inter-observer variability and subjective interpretation. Current artificial intelligence approaches focus on binary classification without revealing tongue features influencing diagnosis, limiting clinical utility. To address these challenges, we propose two key artificial intelligence innovations: (1) Hierarchical Kolmogorov–Arnold Network-based Adaptive Multi-task Experts, a novel multi-task learning architecture with adaptive quota routing that dynamically determines expert allocation based on sample complexity, solving the negative transfer problem; (2) Dynamic Tongue Relationship Graph Convolutional Network, a graph neural network that transforms expert selection patterns into patient-specific relationship graphs, enabling personalized modeling of attribute interactions. These artificial intelligence contributions advance the state-of-the-art in multi-task learning and dynamic graph construction for heterogeneous data, representing our core contributions at the artificial intelligence algorithmic level. As the engineering application of these artificial intelligence innovations, we develop Tongue Network with Dynamic Graph Relationship Learning, a non-invasive and interpretable framework based on tongue images that identifies seven critical tongue attributes and models their interrelationships for liver fibrosis risk screening in Traditional Chinese Medicine-based hepatology practice. Our method achieves 0.751 accuracy and 0.767 area under the curve, with 0.07 and 0.08 improvements over state-of-the-art graph neural networks and multimodal approaches. The framework offers visualizable classifications and relationship structures, providing insights into tongue features and liver fibrosis risk assessment. Codes are available at https://github.com/MedAI-UAIX/TongueNet-DGRL .