MedGraph AI: Agentic Clinical Knowledge Graph

MedGraph AI addresses hospital communication gaps, diagnostic delays, and dangerous drug-drug interactions (DDIs) by integrating neuro-symbolic reasoning into an asynchronous, enterprise-grade clinical command center. Built on AMD ROCm Cloud infrastructure, the architecture leverages Gemma-4-31B-IT for clinical extraction and safety verification alongside DeepSeek-V4.1-Flash for low-latency patient-friendly Generative UI translation. The pipeline orchestrates a 4-node LangGraph workflow: PII Guard Node: Enforces HIPAA compliance by automatically redacting patient identifiers before external inference. Dynamic Graph Reasoning Node: Traverses FalkorDB (medgraph_hospital_brain) across multi-hop relationships (Patient → Medication → Biological Pathway → Clinical Risk → Recommended Alternative → Escalation Specialist). Clinical Verification Gate: Evaluates extracted symptoms, cross-checks contraindications against hospital knowledge graphs, and locks patient discharge if an autonomous confidence threshold (<80%) or severe DDI is detected. Generative UI Node: Translates verified clinical records into 5th-grade reading-level discharge summaries, dynamic medication checklists with safety hold tags, and red-flag warning cards. To close the loop, MedGraph AI pairs PostgreSQL role-based authentication (Doctor vs. Patient portals) with FalkorDB write-backs. When an attending physician reviews and signs off on a staged case, the decision is committed as a permanent episodic memory edge, enabling cross-session auditability. MedGraph AI bridges deterministic graph safety with generative clinical intelligence to protect patient outcomes.