LLM-HFR-RL: Large Language Model (LLM)-Driven Cross-Modal Fine-Grained Alignment and Reinforcement Learning for the Prediction of Heart Failure Risk

ICLR 2026 Conference Submission20319 Authors

Keywords: Large Language Model, Multimodal fusion, Reinforcement Learning

Abstract: Predicting Heart Failure Risk (HFR) using electronic health records (EHR) and generating actionable clinical decisions face significant challenges, including integrating multimodal data, modeling longitudinal temporal patterns, and translating predictions into executable interventions. To address these limitations, this paper proposes the LLM-HFR-RL framework, bridging the gap from risk prediction to clinical decision-making. This framework integrates three key technical innovations: (1) a longitudinal laboratory index summarization method leveraging large language models (LLMs), which transforms discrete test value sequences into clinically meaningful trend summaries; (2) a ternary cross-modal fine-alignment architecture that integrates semantic representations across structured test sequences, LLM-generated trend summaries, and clinical text; and (3) the novel integration of a Reinforcement Learning (RL)-driven decision engine, which learns optimal testing strategies via a multi-objective reward function to dynamically refine clinical decisions. Experimental results demonstrate that LLM-HFR-RL not only significantly improves HFR prediction performance but also forms a high-precision and cost-effective clinical decision support system, providing a new paradigm for intelligent medical intervention.

Primary Area: reinforcement learning

Submission Number: 20319