A DECISION SUPPORT SYSTEM FOR HEART FAILURE MEDICATION

Abstract

Heart failure is a serious long-term condition that affects tens of millions of people worldwide. Clear treatment guidelines exist, yet many patients still do not receive the full set of medicines shown to improve survival and symptoms. This thesis presents a Clinical Decision Support System (CDSS) that helps clinicians apply those guidelines more safely and consistently. The system combines a medical Knowledge Graph, GraphRAG retrieval, deterministic safety rules, and a locally hosted large language model. Medical facts live in a structured graph; retrieval pulls supporting passages from drug labels and guidelines; fixed rules enforce hard safety constraints; the model explains results in Vietnamese or English without owning the final recommendation. An automated pipeline ingests FDA drug labels and heart-failure guidelines into searchable rules, text chunks, and graph links. Clinicians describe a patient in chat and receive guideline-directed therapy suggestions, interaction and dose warnings, and citations for each card. On 50 curated clinical cases, recommendation accuracy reached 94.0%, mean response time was 8.1 seconds, and user satisfaction averaged 4.22 out of 5.0. The evaluation supports hybrid graph-and-rule decision support for supervised heart-failure care, with the clinician retaining final judgment. Hệ hỗ trợ quyết định lâm sàng cho thuốc suy tim, kết hợp đồ thị tri thức, GraphRAG, quy tắc an toàn xác định và LLM chạy cục bộ để đưa ra gợi ý điều trị an toàn, có trích dẫn nguồn, trong đó bác sĩ vẫn giữ quyền quyết định cuối cùng.