[Paper Review] Enhancing Nursing and Elderly Care with Large Language Models: An AI-Driven Framework

Qiao Sun, Jianfei Xie+3 more|arXiv (Cornell University)|Dec 13, 2024

Artificial Intelligence in Healthcare and EducationMedicine3 citations

TL;DR

This paper proposes a specialized large language model (LLM) framework for nursing and elderly care using incremental pre-training and supervised fine-tuning on a novel multilayer Chinese nursing dataset. The system, integrated via LangChain into a dynamic nursing assistant, demonstrates significant performance gains in real-time patient monitoring and personalized interventions, offering a scalable AI-driven solution for aging population care challenges in China.

ABSTRACT

This paper explores the application of large language models (LLMs) in nursing and elderly care, focusing on AI-driven patient monitoring and interaction. We introduce a novel Chinese nursing dataset and implement incremental pre-training (IPT) and supervised fine-tuning (SFT) techniques to enhance LLM performance in specialized tasks. Using LangChain, we develop a dynamic nursing assistant capable of real-time care and personalized interventions. Experimental results demonstrate significant improvements, paving the way for AI-driven solutions to meet the growing demands of healthcare in aging populations.

Motivation & Objective

• Address the growing demand for elderly care in China, where 19.8% of the population is aged 60+, by developing AI-driven solutions to alleviate workforce shortages.
• Overcome the limitations of general-purpose LLMs in nursing by creating a domain-specific model trained on specialized clinical and caregiving data.
• Develop the first multilayer Chinese nursing dataset with text, dialogue, and image-text pairs to support fine-tuning and benchmarking in elderly care.
• Design and implement a LangChain-based AI agent capable of real-time, context-aware nursing assistance with potential for multimodal input integration.
• Ensure ethical compliance through participant consent, data anonymization, and open non-commercial licensing of the dataset.

Proposed method

• Developed a multilayer Chinese nursing dataset comprising 2,510 image-text pairs, single-turn dialogues, multi-turn dialogues, and markdown-formatted clinical text.
• Applied incremental pre-training (IPT) and supervised fine-tuning (SFT) to adapt a foundational LLM to nursing-specific tasks and terminology.
• Integrated the fine-tuned LLM into a LangChain-powered agent architecture to enable dynamic, context-aware interactions and real-time care support.
• Used GPT-4o and InternLM to generate synthetic multi-turn dialogues based on public knowledge, ensuring alignment with clinical guidelines.
• Employed LabelMe for polygon segmentation of people and equipment in image-text pairs, with 10 annotators ensuring data quality and compliance.
• Released the dataset under a CC BY-NC 4.0 license with full attribution and ethical safeguards, including participant consent and data anonymization.

Experimental results

Research questions

• RQ1 Can a fine-tuned LLM significantly improve performance in nursing-specific tasks such as patient monitoring and personalized care planning?
• RQ2 How effective is a multilayer Chinese nursing dataset in enhancing LLM generalization and task-specific accuracy in elderly care?
• RQ3 To what extent can a LangChain-integrated LLM agent support real-time, context-aware nursing interventions?
• RQ4 What are the ethical and practical challenges in collecting and using multimodal nursing data, particularly involving human images and voice?
• RQ5 How does the performance of a domain-adapted LLM compare to general medical LLMs in nursing scenarios requiring continuous monitoring and interaction?

Key findings

• The proposed LLM framework achieved state-of-the-art performance on a newly established benchmark for nursing knowledge and skills, demonstrating superior accuracy in clinical reasoning and care planning.
• Ablation studies confirmed that the multilayer dataset significantly improved model performance, especially in handling complex, context-dependent nursing scenarios.
• The LangChain-integrated agent successfully supported real-time, multi-turn dialogues with personalized care recommendations, showing responsiveness and contextual coherence.
• The dataset’s inclusion of image-text pairs enabled preliminary exploration of multimodal integration, though full visual processing remains a future direction.
• Ethical safeguards, including participant consent, data anonymization, and open non-commercial licensing, were successfully implemented and approved by institutional review boards.
• Despite strong performance, challenges in real-time inference speed, model hallucination, and bias mitigation in sensitive care contexts were identified as key limitations.