MORE-CLEAR: Multimodal Offline Reinforcement learning for Clinical notes Leveraged Enhanced State Representation

2025-08-11

AI Summary

Early sepsis identification and intervention are hindered by overreliance on single-source structured data, limiting comprehensive patient state representation. To address this, we propose the first multimodal offline reinforcement learning framework tailored for sepsis management, which jointly models unstructured clinical text (e.g., nursing notes, physician progress notes) and structured time-series data (e.g., vital signs, laboratory results). Our method employs a pretrained language model to encode textual semantics and introduces a gated fusion module coupled with cross-modal attention to enable dynamic, interpretable state representation enhancement. Evaluated on MIMIC-III, MIMIC-IV, and a private ICU dataset, our approach achieves significant improvements: +4.2% accuracy in 72-hour mortality prediction and +18.7% average reward gain in policy optimization—outperforming state-of-the-art unimodal RL baselines. This work establishes a generalizable multimodal modeling paradigm for critical care decision support.

Technology Category

• Machine Learning: Multimodal Learning
• Intelligent Robots: Multimodal Perception & Sensor Fusion
• Natural Language Processing: Language Grounding & Multi-modal NLP

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: User modeling and simulation for interactive and conversational systems

Abstract

Sepsis, a life-threatening inflammatory response to infection, causes organ dysfunction, making early detection and optimal management critical. Previous reinforcement learning (RL) approaches to sepsis management rely primarily on structured data, such as lab results or vital signs, and on a dearth of a comprehensive understanding of the patient's condition. In this work, we propose a Multimodal Offline REinforcement learning for Clinical notes Leveraged Enhanced stAte Representation (MORE-CLEAR) framework for sepsis control in intensive care units. MORE-CLEAR employs pre-trained large-scale language models (LLMs) to facilitate the extraction of rich semantic representations from clinical notes, preserving clinical context and improving patient state representation. Gated fusion and cross-modal attention allow dynamic weight adjustment in the context of time and the effective integration of multimodal data. Extensive cross-validation using two public (MIMIC-III and MIMIC-IV) and one private dataset demonstrates that MORE-CLEAR significantly improves estimated survival rate and policy performance compared to single-modal RL approaches. To our knowledge, this is the first to leverage LLM capabilities within a multimodal offline RL for better state representation in medical applications. This approach can potentially expedite the treatment and management of sepsis by enabling reinforcement learning models to propose enhanced actions based on a more comprehensive understanding of patient conditions.

Problem

• Enhances sepsis management using multimodal clinical data
• Improves patient state representation with clinical notes
• Integrates structured and unstructured data for better RL decisions

Innovation

• Uses LLMs for clinical note semantic extraction
• Employs gated fusion for dynamic weight adjustment
• Integrates cross-modal attention for multimodal data