An AI-driven LLM agent for analyzing elderly activities of daily living and health data reporting

This project develops an LLM agent to analyze ADL data for elderly individuals, generating insightful health reports. The system processes multimodal sensor data capturing daily activities, identifies patterns, and produces textual wellness summaries. It automates report generation from multidimensional health data (e.g., sleep patterns, heart rate), outputting two formats: a detailed analysis of long/short-term health metric trends with interpretations; and an elderly-optimized bulletin featuring concise conclusions, and intuitive visual symbols for enhanced accessibility. Key challenges included ensuring an accurate interpretation of ADL signals over long periods without losing critical details and generating summarized reports that meet strict structural, visualization and senior-friendly requirements efficiently. To address these, four major contributions were made: 1) Designing a graph memory module to improve stability in long-term health analysis; 2) Constructing a two-level report generation pipeline for end-to-end automated output; 3) Developing a hybrid evaluation framework for systematic quality assessment of generated reports; and 4) Implementing fine-tuning techniques to optimize efficiency and alignment in end-to-end summary generation. Together, these innovations support more reliable, efficient, and personalized health analysis for elderly care.