Sentiment Analysis Using a Large Language Model-Based Approach to Detect Opioids Mixed With Other Substances Via...

Area of Science:

• Public Health
• Data Science
• Computational Linguistics

Background:

• The opioid crisis is a major US health challenge with rising overdose deaths.
• Understanding the epidemic requires enhanced health surveillance.
• Machine learning can analyze social media for opioid user risk identification.

Purpose of the Study:

• To leverage machine learning for analyzing self-reported opioid use patterns from social media.
• To identify opioid use and risk factors non-invasively.
• To improve understanding of opioid effects, especially when mixed with other substances.

Main Methods:

• Analysis of YouTube comments (December 2020 - March 2024) detailing self-reported opioid experiences.
• Manual annotation of comments into categories of positive and negative effects.
• Application of machine learning models, including deep learning, transformer, and large language models (GPT-3.5 Turbo), for sentiment and risk analysis.

Main Results:

• GPT-3.5 Turbo achieved high precision and accuracy in identifying adverse effects and high-risk drug use from YouTube comments.
• The proposed GPT-3.5 Turbo methodology yielded an F1-score of 0.95.
• A 3.26% performance improvement was observed compared to traditional machine learning models (e.g., extreme gradient boosting with F1-score 0.92).

Conclusions:

• Machine learning and large language models can effectively analyze public sentiment on opioid use from social media.
• YouTube comments provide valuable self-reported data on opioid effects and risks.
• The study's methodology enhances overdose risk prediction, improving healthcare responses to the opioid crisis.