ClinNoteAgents: An LLM Multi-Agent System for Predicting and Interpreting Heart Failure 30-Day Readmission from Clinical Notes

Summary

ClinNoteAgents is an LLM-based multi-agent system designed to predict and interpret 30-day heart failure (HF) readmission risk from free-text clinical notes. It transforms unstructured notes into structured clinical/social risk factors and clinician-style abstractions, demonstrating strong performance in risk factor identification and readmission prediction. This approach offers a scalable and interpretable solution, reducing reliance on structured data and manual efforts in healthcare systems.

Medical Relevance

Heart failure readmission imposes a significant clinical and financial burden. This system leverages rich, often underutilized, unstructured clinical data to proactively identify patients at high risk, potentially enabling timely interventions to improve patient outcomes and reduce healthcare costs.

AI Health Application

The paper presents ClinNoteAgents, an LLM multi-agent system designed to predict and interpret 30-day heart failure readmission from unstructured clinical notes. It transforms free-text notes into structured risk factors and clinician-style abstractions to enhance readmission prediction and provide interpretable insights for healthcare providers, thereby serving as a clinical decision support tool.

Key Points

• Addresses the problem of high 30-day heart failure readmission rates and the underutilization of rich, free-text clinical notes for risk analysis due to their unstructured nature.
• Introduces ClinNoteAgents, a novel LLM-based multi-agent framework specifically engineered to process complex, free-text clinical notes.
• Performs a dual transformation: (1) extracting structured representations of clinical and social risk factors for association analysis, and (2) generating clinician-style abstractions for HF 30-day readmission prediction.
• Evaluated on a dataset comprising 3,544 clinical notes from 2,065 patients, where the 30-day HF readmission rate was 35.16%.
• Achieved strong performance in extracting relevant risk factors directly from free-text, identifying key contributing factors, and accurately predicting 30-day HF readmission risk.
• Significantly reduces the reliance on structured EHR fields and minimizes the need for extensive manual annotation and traditional model training.
• Provides a scalable and interpretable approach to note-based HF readmission risk modeling, especially beneficial for data-limited healthcare systems.

Clinical Impact

This system offers a scalable and interpretable solution for HF readmission risk modeling, particularly valuable for healthcare systems with limited structured data. By efficiently converting complex clinical narratives into actionable insights, it can facilitate proactive identification of high-risk patients, guide targeted interventions, and potentially reduce rehospitalization rates, while simultaneously reducing the burden of manual data annotation and model development.

Limitations

Not explicitly mentioned in the abstract. Potential limitations not detailed in the abstract might include generalizability across diverse clinical settings or EHR systems, or specific considerations regarding potential biases inherent in LLMs or their training data.

Future Directions

Not explicitly mentioned in the abstract. Potential future research could involve validating ClinNoteAgents in larger and more diverse patient populations, exploring its integration into real-time clinical decision support systems, or expanding its application to predict readmission for other complex medical conditions.

Medical Domains

Keywords

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.

Comments

10 pages, 2 figures. Submitted to AMIA 2026 Informatics Summit Student Paper Track