RAGNet: Transformer-GNN-Enhanced Cox–Logistic Hybrid Model for Rheumatoid Arthritis Risk Prediction

Rheumatoid Arthritis (RA) is a complex autoimmune disease and early prediction is important for clinical intervention. Nevertheless, existing works have limitations in terms of accuracy and generalization. In this study, we developed an AI-optimized biostatistical model (AIOBM) for accurate prediction risk of RA. The model is a classic Cox proportional hazard model and logistic regression architecture, supplemented by deep learning architecture with a multilayer transformation module, as well as a GNN to model the potential biological correlation of feature network among patients and attention mechanism to select key factors. At the optimization level, we formulated a new learning problem and developed a multi-objective evolutionary algorithm to optimize the joint accuracy and interpretability of a model, encode biostatistical priors as regularization terms into the learning objective so as to drive the model to learn decision paths consistent with medical principles. The whole model is comprised of four modules: feature embedding module (dealing with the multi-modal medical data), cross-representation module (leveraging the Transformer to discover the interaction patterns), GNN inference module, and risk prediction module, constituting a hybrid system with high interpretability and stable predictions. experimental results demonstrate that AIOBM outperforms the state-of-the-art algorithms for real clinical RA data.