대한신장학회

논문분류	춘계학술대회 초록집
제목	Real-Time Intradialytic Hypotension Prediction by Hemodialysis Patient Graph with Large Language Model Summaries
저자	Sanghyuk Kwak
출판정보	2026; 2026(1):
키워드	Hemodialysis, Clinical decision support, Context graph, Retrieval-augmented generation, Large language model
초록	Objectives: To develop a clinical decision support system that represents longitudinal hemodialysis trajectories as a patient–temporal contextual graph, generates evidence-linked large language model (LLM) summaries, and predicts intradialytic hypotension (IDH) with rigorous generalizability assessment. Methods: Twelve months of data (March 2025–February 2026) from 66 patients (9,206 sessions) were analyzed from a hemodialysis management system (Modunuri®), integrating five streams: minute-level machine logs (2.2 million ORU records), vital signs, laboratory results, dry-weight adjustments, and nursing events. Data were transformed into patient-centric contextual graphs (mean 548 nodes per patient), from which time-horizon subgraphs were retrieved for graph-based RAG summaries via a local LLM. For IDH prediction (SBP < 100 mmHg within 30/60 minutes), 39 features were engineered from 5-minute resampled observations (353,590 points). The dataset was split temporally at September 2025 into training and test sets. Robustness was further assessed via LOPO cross-validation and LOPO restricted to SBP > 110 mmHg observations. Results: The retrieval-augmented generation (RAG) pipeline generated multi-horizon summaries with explicit node-level evidence citations. For 30-minute IDH prediction, Random Forest area under the curve–receiver operating characteristic (AUC-ROC) was 0.937 on temporal split, 0.914 with systolic blood pressure (SBP) > 110 mmHg filtering, 0.872 ± 0.117 on leave-one-patient-out (LOPO) cross-validation—where each patient was held out entirely to assess generalizability to unseen individuals—and 0.822 ± 0.130 under the most stringent SBP > 110 mmHg + LOPO evaluation. For 60-minute prediction: 0.921, 0.897, 0.848 ± 0.125, and 0.816 ± 0.119, respectively. Top features were mean arterial pressure (MAP), SBP change from session start, and ultrafiltration rate. Conclusion: A patient-temporal contextual graph provides unified infrastructure for both evidence-linked LLM summaries and IDH prediction, maintaining AUC-ROC above 0.82 even under the most stringent validation, supporting clinical utility for early IDH warning.
원문(PDF)	PDF 원문보기

간행물 검색

대한신장학회 회원관리 규정