PI-EnLLM: Personalized Interactive Healthcare Assistance via Ensembling Large Language Models

Personalized Assistance is critical to maintain an individual’s mental health and well-being. AI-powered interactive agents demonstrate the potential to provide convenient and accessible support to individuals from diverse socio-econo-cultural backgrounds. By facilitating continual monitoring and just-in-time interventions, these agents demonstrate themselves as possible game changers in this application domain. This paper presents PI-EnLLM, a personalized and model-agnostic Large Language Model-based interactive healthcare assistant that can produce a context-aware response to a user query. An effective Cooperative Optimization process leverages the complementary strengths of multiple base LLMs to consent to a response and address a user query. In contrast to performing a tedious, resource-intensive, and task-specific LLM fine-tuning, we draft a context-aware dynamic prompt tuning technique, which can distill useful information into the prompts to generate a personalized response leveraging the user’s unique past conversation context. The complementary knowledge resources of multiple LLMs are utilized to iteratively validate and enhance the response completeness and factuality in parallel. Across two large-scale publicly available datasets and our in-house PsychEd_Care psycho-education Question-Answer (QA) data collection, the proposed PI-EnLLM demonstrates consistent superior performance (e.g., 20-40% improvement in F-measure of the R_L score reported by PI-EnLLM(3-Ensemble) in Psych8K dataset) compared to its individual base LLMs. This proves the effectiveness of the proposed context-aware dynamic prompt tuning toward expediting a cooperative optimization of the generated response via multiple iterations of LLM-specific validation checks. The generated response also reports impressive gain in its factuality and LLM-judge scores, exhibiting enhanced alignment with human preferences. The QA collection in the PsychEd_Care dataset covers essential caregiving topics including Transfer Skills, Nutrition, Dental Care, Bathing and Dressing, Toileting and Incontinence, Behavioral Issues, and Self-Care and will be available to academic researchers in the community after the work is published.