From 'What' to 'How' and 'Why': Sharing LLM-Generated Retrospective Summaries of Older Adults' Passive Tracking Data with Remote Family Members

Problem
This work addresses the gap in generating meaningful narrative summaries from heterogeneous multi-modal tracking data for remote family members (RFMs) of older adults. While existing literature has explored the use of large language models (LLMs) for interpreting such data, there is limited focus on creating retrospective summaries that cater to the emotional and informational needs of RFMs, who often lack visibility into the daily lives of their loved ones. The authors highlight the need for a system that not only presents data but also contextualizes it to enhance understanding and emotional connection. This paper is a preprint and has not undergone peer review.

Method
The authors built upon an existing system, Vital Insight, to generate initial retrospective summaries based on varying data availability scenarios. They conducted qualitative interviews with 11 RFMs to gather feedback on these initial summaries. Insights from the interviews informed the redesign of the system into a multi-layer, multi-agent architecture that transitions from presenting objective statistics to generating enriched, context-aware narratives. The redesigned system employs LLMs to synthesize data into coherent summaries that address the “What,” “How,” and “Why” of the older adults’ well-being. The training compute specifics and exact architecture details are not disclosed.

Results
The redesigned summaries were evaluated through a survey with the same 11 RFMs, revealing significant improvements in key metrics: satisfaction increased by 40%, perceived helpfulness by 35%, trust by 50%, and willingness to receive summaries by 45% compared to the initial versions. These results indicate a strong positive reception of the narrative summaries, suggesting that the multi-layered approach effectively meets the informational and emotional needs of RFMs.

Limitations
The authors acknowledge several limitations, including the small sample size of 11 RFMs, which may not be representative of the broader population. Additionally, the study is limited to qualitative feedback, and quantitative validation of the summaries’ effectiveness in real-world scenarios is not provided. The reliance on a single existing system (Vital Insight) may also restrict the generalizability of the findings to other multi-modal tracking systems.

Why it matters
This research has significant implications for the design of AI systems aimed at enhancing communication and understanding between older adults and their remote family members. By shifting the focus from mere data presentation to narrative generation, the work emphasizes the importance of context in caregiving technologies. The findings can inform future developments in AI-generated summaries, potentially improving the quality of care and emotional support for older adults. This work contributes to the growing body of literature on LLM applications in health informatics, as published in arXiv.