Analysis Model for Infant Incubator Adverse Events Using Retrieval-Augmented Generation Combined With Dual-Adapter Fine-Tuning: Development and Evaluation Study
2026-03-31
Wenke Xia, Wanting Zhu, Tianchun Li, Li Wang, Weiqi Li, Peiming Zhang
Background: Infant incubator adverse events refer to various harmful incidents that occur during the normal use of marketed infant incubators and result in, or may result in, bodily harm. In recent years, however, the number of reported adverse events has continued to rise. This trend has made the monitoring of infant incubator adverse events time-consuming and labor-intensive when relying solely on manual processing by medical device adverse-event monitoring personnel. Meanwhile, general-purpose large language models (LLMs) still face domain knowledge gaps and hallucination issues in specialized fields. Through fine-tuning, LLMs can be adapted to specific application scenarios, while retrieval-augmented generation (RAG) enhances their ability to handle knowledge-intensive tasks. Therefore, LLMs that integrate these 2 technologies hold significant potential for addressing monitoring challenges. Objective: This study aims to enhance the intelligent monitoring of adverse events related to medical devices by integrating RAG with dual-adapter fine-tuning to construct an adverse-event analysis model for infant incubators. The model enables an integrated workflow centered on extraction and analysis, encompassing structured extraction of adverse events, narrative analysis, and regulatory question answering. Methods: This study leveraged adverse-event data from Chinese infant incubators to construct a high-quality dataset through prompt engineering. Technologically, it combined 2 parameter-efficient fine-tuning methods—low-rank adaptation and infused adapter by inhibiting and amplifying inner activations—to achieve efficient adaptation on the Qwen2-7B base model. Simultaneously, it introduced the FINBGE embedding model with supervised contrastive semantic optimization to build a knowledge retrieval system that mitigates hallucination. Results: This study comprises 1565 pediatric disease question-and-answer entries from PediaBench, 2530 specific corpora on adverse events in infant incubators, and 1488 regulatory corpora. Extensive experiments demonstrate the superiority of this analytical model across various metrics. Under the same experimental conditions, the element recall rate reaches 0.815, the accuracy of infant incubator adverse event analysis is 0.898, and the accuracy in regulatory clause question-answering tasks attains 0.938. Conclusions: The analytical model proposed in this study demonstrates significant advantages in analyzing adverse events related to infant incubators, while also achieving substantial improvements in text generation metrics. When combined with RAG, it not only effectively mitigates hallucination but also enhances knowledge timeliness. This study employs the Qwen model as its foundational framework, leveraging large-model fine-tuning and RAG to achieve intelligent analysis of high-risk medical device adverse-event monitoring data. It demonstrates the feasibility of implementing intelligent medical device regulation within the Qwen ecosystem.