A Multi-Agent Framework for Real-Time Sentiment Monitoring and Predictive Analysis of Public Health Policies.

INTRODUCTION

Rapid policy rollouts can trigger localized dissatisfaction that is difficult to detect using text-only monitoring and single-pass large language model pipelines. This study aimed to evaluate whether a multimodal, multi-agent framework improves the accuracy, reliability, and early warning sensitivity of public response surveillance during a long-term care policy monitoring window.

METHODS

This comparative evaluation study analyzed multimodal public discourse captured during a predefined monitoring window by integrating text with images and videos. The sentiment classification outputs were assessed against a human-consensus reference standard using the F1 score. Summarization reliability was quantified as the rate of unverifiable or fabricated claims in the generated policy feedback summaries. Temporal dynamics were characterized using sentiment trajectories, engagement acceleration, and topic subcluster tracking, with policy-relevant drivers estimated as shares of negative discourse volume.

RESULTS

The multi-agent framework achieved a higher sentiment classification performance, with an F1 score of 0.89 compared with 0.82 for a single-pass baseline. Robustness improved most noticeably in sarcastic and implicit complaint content, where negative intent was consistently recovered despite superficially positive phrasing. Generative reliability improved sharply, with unverifiable or fabricated claims decreasing to 1.2% versus 14.0% from the baseline. Multimodal recovery increased the captured discourse volume by 34% and added 4,200 unique data points available only in the images and videos.

CONCLUSION

Multimodal multi-agent monitoring strengthened sentiment validity, reduced summary fabrication, and detected topic-level escalation signals in the observed monitoring window. The framework may support earlier identification of policy implementation issues, but its outputs should be interpreted as decision support signals rather than as substitutes for formal policy evaluation.