Potential of LLM-Generated Lifestyle Adjustment Recommendations Based on Multimodal Data

Keywords: large language models, lifestyle recommendations, multimodal data, personal informatics, digital health, remote work, well-being, self-management, ecological momentary assessment, AI evaluation

TL;DR: This study demonstrates that large language models can transform multimodal self-report and sensor data into safe, relevant, and feasible lifestyle recommendations, offering a proof-of-concept workflow for digital health support.

Abstract: The prevalence of burnout, depression, and stress-related disorders has increased markedly in contemporary societies, particularly in the context of flexible and remote working arrangements. These structural shifts impose novel demands on individuals to self-regulate health, well-being, and productivity—responsibilities that were previously supported by organizational structures in conventional workplaces. Traditional self-management strategies struggle to address the complexity of interacting behavioral, psychological, and physiological determinants. This paper explores the feasibility of employing large language models (LLMs) to generate lifestyle adjustment recommendations based on multimodal data that integrate subjective self-reports with objective sensor-derived measures. To this end, we simulate realistic multimodal time-series data for a prototypical remote worker, design a natural language prompt to elicit recommendations from an LLM, and employ an independent LLM to evaluate the generated outputs in terms of safety, relevance, and feasibility. The results suggest that LLMs are capable of detecting meaningful behavioral patterns and translating them into actionable guidance. This approach has the potential to support individuals in developing adaptive routines for health and productivity management. Future research should emphasize real-world validation, integration with digital health platforms, and the establishment of ethical safeguards.

Submission Number: 132