Mild-stage dementia patients primarily experience two critical symptoms: severe memory loss and emotional instability. Traditional care approaches often struggle to provide consistent, personalized emotional support at scale. A new study introduces DEMENTIA-PLAN, an innovative retrieval-augmented generation (RAG) framework that leverages large language models (LLMs) to enhance conversational support.
Methodology
The system employs a multiple knowledge graph (KG) architecture, integrating two distinct dimensional knowledge representations: a daily routine graph for immediate care requirements and temporal patterns, and a life memory graph for long-term personal histories and emotional connections. A specialized self-reflection planning agent systematically coordinates knowledge retrieval, dynamically evaluating and adjusting the integration of information from these graphs for optimized response generation.
Findings
Through comprehensive LLM-based evaluation and automated metrics, the framework demonstrated significant improvements over baseline models. The integration of the memory KG notably increased empathy and emotional safety scores, validating clinical observations that interacting within familiar contexts stabilizes patient emotions. The planning agent further enhanced problem-solving capabilities, proving highly effective at guiding patients through daily activities.
Impact
DEMENTIA-PLAN represents a significant advancement in the clinical application of AI for dementia care. By bridging the gap between AI tools and caregiver interventions, this technology offers a scalable solution to provide emotionally intelligent and factually accurate support to patients, reducing agitation and improving overall well-being.
Authors: Yutong Song, Chenhan Lyu, Pengfei Zhang, Sabine Brunswicker, Nikil Dutt, Amir Rahmani
This research was powered by the Livi Platform.
Read the full paper: https://arxiv.org/abs/2503.20950?#
Join our network for deeper insights: https://newsletter.livi.health/si
