Fine-Tuning Large Language Models for Effective Nutrition Support in Residential Aged Care: A Domain Expertise...

Area of Science:

• Artificial Intelligence in Healthcare
• Natural Language Processing for Clinical Data

Background:

• Malnutrition is a significant health issue for older adults in residential aged care (RAC).
• Early identification of malnutrition is crucial for timely and effective interventions.
• Transformer-based large language models (LLMs) offer advanced capabilities for clinical tasks.

Purpose of the Study:

• To develop a domain-specific LLM for the RAC setting.
• To leverage LLMs for improved malnutrition identification and prediction in older adults.
• To enhance the application of AI in healthcare for geriatric care.

Main Methods:

• A RoBERTa-based LLM was fine-tuned on 500,000 nursing progress notes from RAC electronic health records (EHRs).
• The model's embeddings were utilized for malnutrition note identification and prediction tasks.
• Performance was evaluated against various established NLP models using 5-fold cross-validation.

Main Results:

• The RAC domain-specific LLM achieved a high F1-score of 0.966 for malnutrition note identification.
• The model demonstrated strong performance in malnutrition prediction, achieving an F1-score of 0.687.
• The developed LLM outperformed baseline and other BERT-based models.

Conclusions:

• Developing specialized LLMs for RAC settings is feasible for identifying and predicting malnutrition.
• This AI-driven approach supports early intervention strategies for malnutrition in older adults.
• Future research will focus on optimizing prediction accuracy and integrating the model into clinical workflows.