Towards Automating Location-Specific Opioid Toxicosurveillance from Twitter via Data Science Methods

Area of Science:

• Public Health
• Computational Social Science
• Data Science

Background:

• Social media platforms offer potential for real-time, population-level monitoring of substance abuse.
• Opioid abuse, encompassing both prescription and illicit substances, remains a significant public health concern.

Purpose of the Study:

• To manually characterize opioid-related tweets.
• To compare abuse/misuse post rates between prescription and illicit opioids.
• To develop and assess machine learning algorithms for automated social media analysis of opioid chatter.

Main Methods:

• Manual annotation of 9006 tweets into four categories.
• Training and comparison of various supervised machine learning algorithms (Deep Convolutional Neural Networks, Support Vector Machines, Random Forests).
• Evaluation of algorithm performance, focusing on accuracy and challenges like lack of context and data imbalance.

Main Results:

• Deep Convolutional Neural Networks achieved the highest accuracy at 70.4%.
• Support Vector Machines and Random Forests showed comparable performance.
• Misclassification was observed due to tweet context limitations and imbalanced datasets.

Conclusions:

• Machine learning algorithms demonstrate potential for automating social media-based opioid abuse monitoring.
• Further improvements are needed to address challenges like data context and imbalance for enhanced accuracy.
• Social media analysis offers a promising avenue for near real-time public health surveillance of opioid trends.