Analysis Model for Infant Incubator Adverse Events Using Retrieval-Augmented Generation Combined With Dual-Adapter

Wenke Xia¹, Wanting Zhu¹, Tianchun Li¹

¹School of Health Science and Engineering, University of Shanghai for Science and Technology, Shanghai, China.

JMIR Medical Informatics

|March 31, 2026

Summary

Related Concept Videos

You might also read

Articles linked to this work by shared authors, journal, and citation graph.

Sort by

Same author

AI-Driven Medical Device Risk Management: A New Paradigm Integrating Large Language Models and Prompt Engineering for Standard-Risk Knowledge Graph Construction and Application.

Risk management and healthcare policy·2026

See all related articles

This study introduces an intelligent model for infant incubator adverse event monitoring, combining retrieval-augmented generation (RAG) and fine-tuning to improve accuracy and reduce manual workload in medical device safety surveillance.

Area of Science:

Medical Device Safety
Artificial Intelligence in Healthcare
Clinical Engineering

Background:

Rising infant incubator adverse events necessitate efficient monitoring solutions.
Manual processing of adverse events is labor-intensive and time-consuming.
General-purpose large language models (LLMs) have limitations in specialized medical domains.

Purpose of the Study:

To develop an intelligent adverse event analysis model for infant incubators.
To integrate retrieval-augmented generation (RAG) with dual-adapter fine-tuning for enhanced monitoring.
To create a workflow for structured extraction, narrative analysis, and regulatory question answering.

Main Methods:

Constructed a high-quality dataset using prompt engineering with Chinese infant incubator adverse-event data.

Keywords:

adverse events in medical devices fine-tuning techniques infant incubators large language models retrieval-augmented generation

Employed parameter-efficient fine-tuning (low-rank adaptation and infused adapter) on the Qwen2-7B model.

Integrated the FINBGE embedding model with supervised contrastive semantic optimization for knowledge retrieval.

Main Results:

Achieved an element recall rate of 0.815.
Demonstrated an accuracy of 0.898 in infant incubator adverse event analysis.
Attained an accuracy of 0.938 in regulatory clause question-answering tasks.

Conclusions:

The proposed analytical model significantly improves infant incubator adverse event analysis and text generation.
RAG integration effectively mitigates hallucination and enhances knowledge timeliness.
Demonstrated the feasibility of intelligent medical device regulation within the Qwen ecosystem.

Related Experiment Video

Analysis Model for Infant Incubator Adverse Events Using Retrieval-Augmented Generation Combined With Dual-Adapter...