Personalized Remote Health Monitoring and Clinical Report Generation Using Multi-Agent LLM Framework on Wearable Data

Keywords: AI in healthcare, LLM, Agetentic AI, wearables

TL;DR: We build a modular multi-agent LLM that turns wearable time-series and personal metadata into clinician-style daily/weekly reports. Validated on a 50-user synthetic cohort, it shows strong anomaly detection and positive expert reviews.

Abstract:

Remote health monitoring with consumer wearables can continuously capture sleep, activity, and cardiorespiratory signals, but raw metrics are difficult to translate into clinically meaningful guidance. We present a modular multi-agent large language model (LLM) framework that converts longitudinal wearable data plus personal metadata (age, BMI, smoking status, and comorbidities) into clinician-style daily and weekly narratives with targeted recommendations. Our pipeline comprises specialized agents for (i) anomaly detection, (ii) trend summarization, (iii) lifestyle suggestions, (iv) final daily synthesis, and (v) weekly aggregation. To enable reproducible evaluation without PHI, we generate a medically grounded synthetic dataset of \emph{50} users over \emph{7} days, with anomalies injected in \emph{$35\%$} of users. Each user contributes $2016$ five-minute physiological samples, $1008$ ten-minute step intervals, and $224$ sleep stage entries per week, yielding more than $160{,}000$ structured records across the cohort.

On injected anomalous events, the Anomaly Agent achieves precision $=0.89$ and recall $=0.91$. In a blinded expert review (two clinicians, one public-health researcher) of system outputs, average ratings (scale $1 \text{--} 3$) were: clinical relevance $=2.7$, personalization $=2.5$, and clarity $=2.8$. An ablation shows that removing personalization reduces individualization by approximately $\sim 24\% \; (\approx 25\%)$, and replacing the multi-agent pipeline with a single-agent baseline lowers average scores across relevance, personalization, and clarity.

\textbf{Code \& data:} \url{https://github.com/AliArshadswl/Agents4Science}.

Submission Number: 184