클린노트에이전트 대형언어모델 기반 다중‑에이전트 프레임워크를 활용한 심부전 30일 재입원 위험 예측

📝 Abstract

Heart failure (HF) is one of the leading causes of rehospitalization among older adults in the United States. Although clinical notes contain rich, detailed patient information and make up a large portion of electronic health records (EHRs), they remain underutilized for HF readmission risk analysis. Traditional computational models for HF readmission often rely on expert-crafted rules, medical thesauri, and ontologies to interpret clinical notes, which are typically written under time pressure and may contain misspellings, abbreviations, and domain-specific jargon. We present ClinNoteAgents, an LLM-based multi-agent framework that transforms free-text clinical notes into (1) structured representations of clinical and social risk factors for association analysis and (2) clinician-style abstractions for HF 30-day readmission prediction. We evaluate ClinNoteAgents on 3,544 notes from 2,065 patients (readmission rate=35.16%), demonstrating strong performance in extracting risk factors from free-text, identifying key contributing factors, and predicting readmission risk. By reducing reliance on structured fields and minimizing manual annotation and model training, ClinNoteAgents provides a scalable and interpretable approach to note-based HF readmission risk modeling in data-limited healthcare systems.

💡 Analysis

Heart failure (HF) is one of the leading causes of rehospitalization among older adults in the United States. Although clinical notes contain rich, detailed patient information and make up a large portion of electronic health records (EHRs), they remain underutilized for HF readmission risk analysis. Traditional computational models for HF readmission often rely on expert-crafted rules, medical thesauri, and ontologies to interpret clinical notes, which are typically written under time pressure and may contain misspellings, abbreviations, and domain-specific jargon. We present ClinNoteAgents, an LLM-based multi-agent framework that transforms free-text clinical notes into (1) structured representations of clinical and social risk factors for association analysis and (2) clinician-style abstractions for HF 30-day readmission prediction. We evaluate ClinNoteAgents on 3,544 notes from 2,065 patients (readmission rate=35.16%), demonstrating strong performance in extracting risk factors from free-text, identifying key contributing factors, and predicting readmission risk. By reducing reliance on structured fields and minimizing manual annotation and model training, ClinNoteAgents provides a scalable and interpretable approach to note-based HF readmission risk modeling in data-limited healthcare systems.

📄 Content

ClinNoteAgents: An LLM Multi-Agent System for Predicting and Interpreting Heart Failure 30-Day Readmission from Clinical Notes Rongjia Zhou1, Chengzhuo Li1, Carl Yang, PhD1, Jiaying Lu, PhD1 1Emory University, Atlanta, GA, USA ABSTRACT Heart failure (HF) is one of the leading causes of rehospitalization among older adults in the United States. Al- though clinical notes contain rich, detailed patient information and make up a large portion of electronic health records (EHRs), they remain underutilized for HF readmission risk analysis. Traditional computational models for HF readmission often rely on expert-crafted rules, medical thesauri, and ontologies to interpret clinical notes, which are typically written under time pressure and may contain misspellings, abbreviations, and domain-specific jargon. We present ClinNoteAgents, an LLM-based multi-agent framework that transforms free-text clinical notes into (1) struc- tured representations of clinical and social risk factors for association analysis and (2) clinician-style abstractions for HF 30-day readmission prediction. We evaluate ClinNoteAgents on 3,544 notes from 2,065 patients (readmission rate=35.16%), demonstrating strong performance in extracting risk factors from free-text, identifying key contributing factors, and predicting readmission risk. By reducing reliance on structured fields and minimizing manual annotation and model training, ClinNoteAgents provides a scalable and interpretable approach to note-based HF readmission risk modeling in data-limited healthcare systems. INTRODUCTION Heart failure (HF) remains a major global health challenge, affecting more than 55 million individuals worldwide, with nearly 80% of cardiovascular deaths occurring in low- and middle-income nations.1 Approximately 25% of HF pa- tients are readmitted within 30 days,2 imposing substantial clinical and financial burdens on health systems. Multiple studies have identified diverse contributors to 30-day readmission of HF, including HF exacerbation,3 comorbidities such as chronic obstructive pulmonary disease, chronic kidney disease, anemia infection,4 and socioeconomic factors.5 Given the complex, multi-factorial nature of HF, computational risk modeling for HF readmission often requires com- prehensive longitudinal patient data.6,7 Electronic health record (EHR) data, which capture demographics, diagnoses, laboratory results, and medications, have therefore become a central data source for studying HF outcomes, including 30-day readmission.8,9 While containing valuable multimodal health information for HF patients, EHRs are often incomplete or unavailable in developing countries10 due to financial, technological, and organizational barriers.11 This challenge is particularly evident in developing countries in Asia and Africa. Many hospitals in Bangladesh and In- donesia continue to rely on handwritten or locally stored digital notes due to financial constraints, limited IT capacity, and poor interoperability infrastructure.12,13 Similarly, healthcare facilities in Kenya, Uganda, and Ghana frequently experience unstable internet connectivity and shortages of trained health informatics personnel.14,15 In these settings, unstructured clinical notes often serve as the primary source of documented patient information. Despite widespread adoption of EHR in the U.S., around 80% of clinical information remains embedded in free-text notes.16 Therefore, clinical notes offer a pragmatic and scalable strategy for constructing predictive models when access to structured EHR data is limited. Early HF readmission models relied primarily on structured EHR data such as demographics, socioeconomic status, medical history, and laboratory measurements.17,18 More recent work has incorporated unstructured clinical notes to capture richer contextual and temporal information,19 with hybrid models combining structured variables and note- derived embeddings yielding further gains.20 This shift underscores the potential of text-based modeling for early detection of readmission risk. Growing evidence also highlights the importance of social determinants of health (SDOHs), defined by the World Health Organization as the conditions in which individuals live and work.21 SDOHs have been repeatedly linked to HF outcomes, with employment status, housing stability, and social support identified as major contributors to readmission risk.22 However, SDOHs are rarely structured in EHRs, limiting their use in predictive systems. Recent advances in natural language processing (NLP) and large language models (LLMs) have enabled the extraction of both clinical and social risk factors from unstructured notes, with machine learning and arXiv:2512.07081v1 [cs.AI] 8 Dec 2025 transformer models achieving state-of-the-art performance in HF readmission prediction using discharge notes.23,24 Summarization-based preprocessing is shown to further enhance predictive signal quality,25 and recent studies show that LLMs c

This content is AI-processed based on ArXiv data.