CLOSED DOMAIN QUESTION ANSWERING WITH LANGUAGE MODELS: APPLICATION OF RETRIEVAL-AUGMENTED GENERATION AND PARAMETER EFFICIENT FINE-TUNING IN HEALTHCARE

Dementia care presents significant challenges for informal caregivers, particularly in managing behavioral symptoms that affect over 90% of individuals with Alzheimer’s Disease and Related Dementias (ADRD) during the moderate-to-severe stages. These symptoms, including agitation, wandering, and repetitive activities, impose emotional and physical burdens on caregivers, often exacerbated by a lack of reliable, accessible, and personalized resources. Non-pharmacological interventions, while evidence-based, are underutilized due to knowledge gaps and the inefficiency of traditional training and information retrieval methods. This research explores the adaptation of large language models (LLMs) to address these challenges by developing a framework for closed-domain Question Answering (QA) systems, with a focus on dementia care practices. The study employs a hybrid methodology combining Retrieval-Augmented Generation (RAG) with Parameter-Efficient Fine-Tuning (PEFT) to create a scalable, efficient, and domain-specific solution. By leveraging structured datasets of dementia-related question-answer pairs and integrating advanced techniques like context injection and model compression, the system dynamically generates accurate, actionable, and contextually relevant responses. The framework is evaluated using standard metrics such as BLEU and ROUGE. Results demonstrate significant improvements in relevance and precision over baseline models, underscoring the effectiveness of RAG and PEFT in domain-specific adaptations of LLMs. A case study featuring the AI-DECAVA app exemplifies the framework's practical application, offering personalized, real-time support to caregivers. Co-created with dementia care experts and informal caregivers, the app bridges critical knowledge gaps, enhances caregiving capabilities, and reduces caregiver stress through evidence-based, non-pharmacological interventions. This work highlights the transformative potential of LLMs in specialized domains, demonstrating their ability to improve real-world outcomes in dementia care. Beyond healthcare, the proposed framework serves as a model for adapting AI to address the unique needs of other domains, paving the way for scalable, efficient, and impactful applications of artificial intelligence.