Background: Social isolation is recognized as a major modifiable risk factor for dementia. The Internet-Based Conversational Clinical Trial (I-CONECT, NCT02871921) demonstrated that semi-structured engaging conversations can effectively reduce social isolation and slow cognitive decline among older adults with mild cognitive impairment (MCI). However, the scalability of this intervention is constrained by the availability and cost of trained human interviewers, limiting its broader dissemination and long-term sustainability.

 

Methods: To address this challenge, we developed the AI-CONECT chatbot, a Responsible-AI system designed to implement the I-CONECT conversational intervention protocol using large language models (LLMs) and optimized prompts. This approach enables accessible and affordable intervention without the human resource limitations. We trained the prompts of the AI-CONECT chatbot using transcribed conversations from the I-CONECT study via a privacy-preserving learning framework and evaluated it with a novel strategy using "virtual users" (digital twins) - LLM replicas of participants from the I-CONECT clinical trial. We compared the engagement levels of virtual users with ChatGPT and the AI-CONECT chatbot, namely, ChatGPT with trained prompts.

 

Results: We simulated conversations between chatbots and the virtual users - 9 with MCI and 9 with normal cognition (NC). Each virtual user completed 60 virtual conversations (with <=40 turns, approximately 3000-4000 words or 20-30 minutes per session) with ChatGPT or AI-CONECT chatbots (ChatGPT with trained prompts). We found that training chatbots' prompts on the I-CONECT conversations significantly improved AI interviewers' capability in engaging virtual users, measured by the user's word ratio (% of words provided by users out of total words during each session). The higher word ratio implies better user conversational engagement. The improvement is significant both in MCI (ChatGPT: mean=0.235; AI-CONECT: mean=0.489; p=0.0004 by Mann-Whitney U test) and NC (ChatGPT: mean=0.365; AI-CONECT: mean=0.626; p=0.002) virtual users, and approaches human interviewers' performance (with MCI: mean=0.570; with NC: mean=0.665).

 

Conclusions: These findings demonstrate that the AI-CONECT chatbot can engage older adults in cognitively stimulating dialogues in simulated conversations. Our preliminary results demonstrate a pathway for integrating Responsible AI to offer a practical and scalable alternative to human-facilitated I-CONECT interventions toward delaying cognitive decline.