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

Area of Science:

• Computational Social Science
• Artificial Intelligence in Public Policy
• Natural Language Processing and Computer Vision Integration

Background:

• Rapid policy rollouts often lead to localized public dissatisfaction.
• Traditional text-only monitoring and single-pass large language models (LLMs) struggle to detect subtle or nuanced public sentiment.
• Existing methods lack the sensitivity for early warning of policy implementation issues.

Purpose of the Study:

• To evaluate a multimodal, multi-agent framework for public response surveillance.
• To assess improvements in accuracy, reliability, and early warning sensitivity compared to baseline methods.
• To monitor public discourse during a long-term care policy implementation window.

Main Methods:

• Comparative evaluation of a multimodal, multi-agent framework against a single-pass baseline.
• Integration of text, image, and video data for comprehensive public discourse analysis.
• Sentiment classification assessed via F1 score against human consensus; summarization reliability measured by fabricated claims rate; temporal dynamics analyzed through sentiment trajectories and topic tracking.

Main Results:

• The multimodal multi-agent framework achieved superior sentiment classification (F1 score 0.89 vs. 0.82 baseline), particularly for nuanced content like sarcasm.
• Generative reliability significantly improved, reducing fabricated claims from 14.0% to 1.2%.
• Multimodal data integration increased discourse volume by 34%, adding 4,200 unique data points from visual media.

Conclusions:

• The multimodal multi-agent monitoring framework enhances sentiment validity and reduces summary fabrication.
• The system effectively detects topic-level escalation signals, supporting earlier identification of policy implementation challenges.
• Outputs serve as valuable decision support signals, complementing formal policy evaluation.