AI-enabled Clinical Decision Support System (CDSS) for Common Non-communicable Diseases

Translating complex, multi-page clinical guidelines into reliable digital decision pathways remains a major challenge in clinical decision support. Standard one-shot Large Language Models (LLMs) risk hallucinations and lack error traceability, undermining clinicians' trust in automated recommendations. To address this gap, we introduce CDTagent, an explainable multi-agent orchestration architecture designed to convert dense medical guidelines into auditable, executable Clinical Decision Forests. 



Rather than relying on ungrounded end-to-end generation, CDTagent enforces a single-writer controller architecture where worker agents draft decision trees that must satisfy six zero-tolerance gates (Coverage, Source Support, Structure, Handoff, Topology, and Dependency Closure) before canonical promotion. At runtime, execution is strictly deterministic, while LLMs operate strictly as advisory tools. 

Evaluated on Vietnam’s 77-page national Type-2 Diabetes guideline (Decision 5481/QĐ-BYT), CDTagent successfully constructed 21 fully-cited trees spanning 498 nodes with 100% citation fidelity (348/348 steps) directly anchored to source text. In clinical question-answering, the system achieved a 97.97% strict grounded rate and successfully abstained on 10/10 ungrounded queries. CDTagent provides a transparent, zero-hallucination clinical assistant that preserves provenance for human-in-the-loop validation.