Rule-augmented LLM framework for detecting unreasonableness in ICU

Abstract

This paper proposes a rule-augmented model system for detecting unreasonable activities in Intensive Care Unit (ICU) hospitalization, mainly leveraging a large language model (LLM). The system is built on DeepSeek-R1-32B and integrates existing unreasonable activities in ICU hospitalization into health insurance systems through prompt learning techniques. Compared to traditional fixed-threshold rules, the large model augmented with rules possesses the ability to identify errors and exhibits a certain degree of emergent capabilities. In addition, it provides detailed and interpretable explanations for detected unreasonableness, helping the health insurance fund supervision perform efficient and accurate reviews. The framework includes two main sub-models: a discriminator for rule judgment, and an evaluator accuracy enhancement. Training data were derived from anonymized records from multiple hospitals and pre-processed to form the first domestic dataset tailored to unreasonable ICU billing detection tasks. The experimental results validate the effectiveness and practical value of the proposed system.