RSA-KG: A graph-based rag enhanced AI knowledge graph for recurrent spontaneous abortions diagnosis and clinical decision support

Recurrent spontaneous abortion (RSA), or recurrent pregnancy loss (RPL), affects 1%-5% of reproductive-aged women worldwide. Its multifactorial etiology, involving complex genetic, immunologic, and endocrine factors, makes diagnosis challenging. Current clinical guidelines are limited, especially in addressing idiopathic cases (which account for 40% of patients) and integrating emerging biomarkers.

While AI has potential, conventional models and even standard retrieval-augmented generation (RAG) systems are constrained . They often fail to capture systemic pathophysiological networks, adapt poorly to guideline updates, and inefficiently fuse structured (e.g., tabular lab data) and unstructured (e.g., imaging reports) data.

To overcome this, the authors developed RSA-KG, a system built on a GraphRAG architecture to dynamically integrate multimodal data into a domain-specific knowledge graph.

Material and Methods

The system's knowledge base was built by collecting and processing data from 5 major international RSA guidelines, including those from RCOG (UK), CSOG (China), ESHRE (Europe), ASRM (USA), and DGGG (Germany/Austria/Switzerland). Data also included literature, clinical cases, and molecular biomarker data. Multimodal models like SLANet-plus and Doubao Vision were used to extract and process text, tables, and figures from the source documents.

The RSA-KG knowledge graph was constructed using the LightRAG model. A sophisticated multi-model consensus mechanism was employed. A diverse panel of high-performance LLMs (including Doubao 1.5 Pro, Llama-3.3-70B, Qwen 2.5 Max, DeepSeek V3, Grok-3, Claude 3.7, Gemini 2.0 flash, and Chatgpt-03) processed knowledge "chunks" in parallel to identify entities (e.g., "etiological factors") and relationships (e.g., "diagnostic guidelines") .

A knowledge triplet (entity-relationship-entity) was only added to the graph if it surpassed a high consensus threshold among the models. Triplets with low consensus or conflicts were flagged for manual adjudication by a panel of senior reproductive medicine specialists.

A comprehensive evaluation dataset of 3,000 specialized questions was meticulously developed. Questions were stratified into three difficulty levels: Level 1 (foundational knowledge, 40%), Level 2 (clinical application, 40%), and Level 3 (complex clinical reasoning, 20%) .

A multi-disciplinary expert panel (including reproductive specialists, clinical geneticists, and immunologists) validated the questions and "gold standard" answers through a double-blind review and consensus protocol . For real-world validation, 10 clinicians (5 mid-level, 5 senior-level) specializing in reproductive medicine were invited to score the model's outputs on clinical cases using a 1-10 point scale based on accuracy, relevance, practicality, and other criteria.

Results

In the 3,000-question evaluation, RSA-KG-enhanced models consistently and significantly outperformed both Naive_RAG-enhanced models and the raw models (p<0.01). The best-performing model, Deepseek-R1, achieved an accuracy of 86.5% with RSA-KG enhancement, compared to 82.4% with Naive_RAG and 76.5% as a raw model20. Models like Doubao, Grok-3, and Claude-3.7 also showed significant accuracy improvements with RSA-KG.

The RSA-KG-enhanced models also substantially outperformed specialized medical LLMs. For instance, the RSA-KG enhanced DeepSeek-R1 (86.5%) and Doubao-1.5Pro (83.9%) were far more accurate than the medical LLM WisediagTk (64.6%) and Wisediag (64.3%) on the same dataset.

In the qualitative assessment of clinical cases, the outputs from the RSA-KG enhanced model received significantly higher scores from the 10 clinicians compared to the raw models. For example, the RSA-KG-enhanced DeepSeek-R1 received a composite score of 91, while its raw counterpart scored only 80. Similarly, RSA-KG Chatgpt-03 scored 89, versus 78 for the raw model.

A case study for Antiphospholipid Syndrome (APS) (Figure 4) visually demonstrated that the RSA-KG-enhanced Deepseek R1 provided a more comprehensive, accurate, and guideline-sourced diagnosis, treatment plan, and prognosis compared to the raw model or WisediagTK .

Discussion and Conclusion

This study successfully developed and validated RSA-KG, a novel graph-enhanced AI framework for clinical decision support in RSA. The system's strength lies in its ability to handle complex, multifactorial cases where raw models and naive RAG systems fail.

By structuring knowledge, RSA-KG enables "multihop reasoning" that mimics expert clinical thought, allowing it to excel in complex diagnostic scenarios like APS.

Key limitations are acknowledged, including potential bias from the selected guidelines, an evaluation dataset that excludes the very latest biomarkers (discovered after 2025), and an expert validation panel limited to a single discipline (reproductive medicine) .

Future work must focus on a rigorous pathway to clinical adoption, including multicenter trials for broader validation, seamless integration with hospital EHR systems, and adherence to privacy-preserving standards like HIPAA and GDPR . By addressing these challenges, RSA-KG can serve as a foundational framework for advancing AI-assisted management in RSA and other complex reproductive health disorders.