ClinRAG-GRAPH: Clinical-prior Retrieval-Augmented Graph Model with Domain Adversarial Learning for Breast pCR Prediction

Technical Summary: ClinRAG-GRAPH

Problem Statement
Accurate prediction of pathological complete response (pCR) to neoadjuvant chemotherapy (NAC) in breast cancer prior to treatment is critical for treatment stratification. However, robust pre-treatment prediction remains challenging due to three primary limitations in existing approaches: (1) insufficient cross-modal modeling, where simple fusion operators fail to exploit structured dependencies between imaging, clinical, and pathological data; (2) multicenter imaging heterogeneity, where scanner and protocol differences introduce domain-specific biases that hinder cross-center generalization; and (3) a lack of evidence-grounded interpretability, as most models output scalar probabilities without retrieving case-based evidence to support clinical decision-making.

Methodology: ClinRAG-GRAPH
The authors propose ClinRAG-GRAPH, a Clinically informed Retrieval-Augmented Generation Graph framework designed to integrate DCE-MRI, structured clinical variables, and biopsy-derived pathological biomarkers. The architecture consists of four core components:

1. Hierarchical Semantic Clinical-prior Graph Construction:
   Instead of homogeneous connectivity, the model constructs an intra-patient directed graph where nodes represent imaging and clinical/pathological variables. Edges are categorized into three semantic tiers based on clinical guidelines and radiologist consensus:

   • Strong prior: Fixed, high-confidence associations (e.g., molecular biomarker couplings).
   • Soft prior: Clinically agreed relations with inter-patient variability.
   • Learnable prior: Uncertain relations initialized with low weight and learned from data.
     These priors are injected as log-biases on attention logits only at the first graph layer to guide early cross-modal information flow without destabilizing optimization.

2. PRAttn-RGCN (Prior-guided Relational Attentional R-GCN):
   A Relation-Aware Graph Convolutional Network utilizes the constructed graph to fuse heterogeneous modalities. It employs relation-specific attention mechanisms where the prior strength guides the initial message passing. The model aggregates node embeddings via mean pooling to generate a patient representation for pCR prediction.

3. Domain-Adversarial Learning:
   To address multicenter MRI heterogeneity, a dual-branch adversarial strategy is employed. A gradient-reversal layer (GRL) connects the imaging encoder to a domain classifier. This forces the imaging representation to be discriminative for pCR while suppressing scanner/protocol-specific cues, thereby improving domain invariance.

4. LLM-driven Subgraph RAG Module:
   To enhance interpretability and robustness under distribution shifts, the framework incorporates a Large Language Model (LLM) retriever (DeepSeek-V3.2). During inference, the system retrieves clinically analogous historical cases based on learned graph embeddings and attention signatures. The retrieved evidence is fused with the backbone prediction to provide an evidence-grounded inference. The LLM also generates structured, verifiable post-hoc explanations based on the retrieval metadata and driver-edge cues.

Key Contributions

• Graph Construction: The introduction of a hierarchical clinical-prior graph that encodes medical knowledge directly into the edge structure, avoiding the need for auxiliary loss constraints.
• Robustness Mechanism: The design of a prior-guided relational attention mechanism combined with adversarial learning to mitigate multicenter MRI protocol bias while preserving task-relevant features.
• Interpretability: The integration of an LLM-driven subgraph RAG module that retrieves similar cases and provides structured, evidence-based rationales for predictions, moving beyond scalar probability outputs.
• Validation: The assembly and evaluation of a large-scale, multicenter NAC breast cancer cohort comprising two public datasets and three in-house centers.

Results
The model was evaluated on internal test sets (aggregated from three centers) and two external test sets (Ycenter and Qcenter).

• Performance: ClinRAG-GRAPH achieved an AUC of 0.815 (95% CI: 0.738–0.885) on the internal test set, outperforming state-of-the-art baselines including R-GCN (0.706), LMF (0.678), and iMRhpc (0.699).
• Generalization: On external test sets, the model maintained robust performance with AUCs of 0.774 (Ycenter) and 0.712 (Qcenter). Notably, it achieved the highest specificity (0.946) on Ycenter and the best balanced accuracy (0.676) on the challenging Qcenter cohort.
• Ablation: Experiments confirmed that each component contributes to performance. The clinical-prior graph provided the initial gain over a standard R-GCN, while adversarial decoupling and the RAG module further improved generalization on external domains.
• Interpretability: SHAP analysis demonstrated that the model captures clinically meaningful cross-modal dependencies (e.g., ER, PR, HER2, Ki67, age, and MRI features) that vary across molecular subtypes, validating the structured relational reasoning.

Significance
The paper claims that ClinRAG-GRAPH offers a significant advancement in pre-treatment pCR prediction by addressing the specific challenges of multimodal integration, domain shift, and interpretability. By combining structured clinical priors with adversarial domain adaptation and retrieval-augmented generation, the framework demonstrates competitive and consistent performance across diverse centers. The authors posit that this approach supports individualized NAC treatment strategies, particularly for early escalation or de-escalation, by providing predictions that are not only accurate but also grounded in retrievable clinical evidence. The code is made available to facilitate further research in this domain.