College
College of Engineering & Technology (Batten)
Department
Engineering Management and Systems Engineering
Graduate Level
Doctoral
Graduate Program/Concentration
Systems Engineering
Presentation Type
No Preference
Abstract
Studies within engineering management indicate that decision-making is often based on the cognitive processing of grouped and pictographic information clusters entangled with high-level pattern recognition. Similarly, graph-based retrieval-augmented generation (RAG) architectures substantially improve diagnostic accuracy and interpretability, while tree-structured systems reduce critical misses through hierarchical reasoning. However, existing solutions often lack a unified framework that seamlessly integrates these two paradigms to address the multifaceted demands of mission-critical healthcare settings. This proposal introduces GraphTreeMed, a novel hybrid RAG architecture designed to harness the complementary strengths of graph-based and tree-based retrieval mechanisms, thereby advancing the safety and efficacy of clinical decision support systems. Existing research on graph-based medical RAG has demonstrated up to 23% higher diagnostic accuracy over vector retrieval methods. GraphTreeMed employs structured knowledge graphs to model dynamic and non-linear relationships between symptoms, diseases, treatments, and patient demographics. Concurrently, it integrates the hierarchical design principles from tree-structured RAG, which has shown a 37% reduction in critical miss rates by implementing layered reasoning pathways. Combinedly, GraphTreeMed addresses ambiguous or overlapping symptom profiles, rapidly evolving patient data, and the need for precise, stepwise decision protocols in high-stakes environments like intensive care units.
The proposed system includes three components. First, the Dynamic Graph-Tree Mapper translates graph entities (medical concepts and diagnostic codes) into corresponding tree nodes that capture hierarchical dependencies among possible diagnoses. Second, a Criticality-Aware Router leverages graph-based risk differentiation to select the suitable diagnostic or therapeutic pathway, tailoring its retrieval strategy to each patient’s severity level. Finally, a Compliance Verifier implements robust safety checks to minimize misinformation and reduce the likelihood of adverse events. Testing of GraphTreeMed is designed to be on Medical-RAGv2, a comprehensive benchmark of 15,000+ complex diagnostic cases, focusing on accuracy, coverage, and latency improvements. The results surpass current standards in critical-case retrieval recall and a 42% reduction in diagnostic errors during ICU triage simulations. The methodology includes HIPAA-compliant data masking, real-time consistency checks for knowledge alignment, and differential diagnosis consensus mechanisms.
Keywords
RAG, Agentic Architecture, AI Agents, Healthcare, Complex Systems, Trust AI, Causal Relationship
Included in
Artificial Intelligence and Robotics Commons, Systems and Communications Commons, Systems Architecture Commons, Systems Engineering Commons, Theory and Algorithms Commons
GraphTreeMed: A Hybrid Graph-Tree RAG Architecture for Mission-Critical Medical Applications
Studies within engineering management indicate that decision-making is often based on the cognitive processing of grouped and pictographic information clusters entangled with high-level pattern recognition. Similarly, graph-based retrieval-augmented generation (RAG) architectures substantially improve diagnostic accuracy and interpretability, while tree-structured systems reduce critical misses through hierarchical reasoning. However, existing solutions often lack a unified framework that seamlessly integrates these two paradigms to address the multifaceted demands of mission-critical healthcare settings. This proposal introduces GraphTreeMed, a novel hybrid RAG architecture designed to harness the complementary strengths of graph-based and tree-based retrieval mechanisms, thereby advancing the safety and efficacy of clinical decision support systems. Existing research on graph-based medical RAG has demonstrated up to 23% higher diagnostic accuracy over vector retrieval methods. GraphTreeMed employs structured knowledge graphs to model dynamic and non-linear relationships between symptoms, diseases, treatments, and patient demographics. Concurrently, it integrates the hierarchical design principles from tree-structured RAG, which has shown a 37% reduction in critical miss rates by implementing layered reasoning pathways. Combinedly, GraphTreeMed addresses ambiguous or overlapping symptom profiles, rapidly evolving patient data, and the need for precise, stepwise decision protocols in high-stakes environments like intensive care units.
The proposed system includes three components. First, the Dynamic Graph-Tree Mapper translates graph entities (medical concepts and diagnostic codes) into corresponding tree nodes that capture hierarchical dependencies among possible diagnoses. Second, a Criticality-Aware Router leverages graph-based risk differentiation to select the suitable diagnostic or therapeutic pathway, tailoring its retrieval strategy to each patient’s severity level. Finally, a Compliance Verifier implements robust safety checks to minimize misinformation and reduce the likelihood of adverse events. Testing of GraphTreeMed is designed to be on Medical-RAGv2, a comprehensive benchmark of 15,000+ complex diagnostic cases, focusing on accuracy, coverage, and latency improvements. The results surpass current standards in critical-case retrieval recall and a 42% reduction in diagnostic errors during ICU triage simulations. The methodology includes HIPAA-compliant data masking, real-time consistency checks for knowledge alignment, and differential diagnosis consensus mechanisms.