A machine learning based framework for identifying consumer product injuries from social media data.

Area of Science:

• Consumer Product Safety
• Public Health Surveillance
• Machine Learning Applications

Background:

• Millions of product-related injuries occur annually, with traditional surveillance methods causing delays in identifying injury patterns.
• Current methods, relying on hospital data, are slow, hindering timely preventive actions like product recalls.
• This delay leads to continued consumer product-related injuries.

Purpose of the Study:

• To develop and evaluate a machine learning (ML) framework for real-time injury surveillance.
• To extract product injury details from social media for rapid trend identification and intervention.
• To improve the speed and efficacy of consumer product safety monitoring.

Main Methods:

• A two-stage ML framework was proposed, utilizing social media posts (Reddit) and National Electronic Injury Surveillance System (NEISS) data.
• Stage 1 classified posts as injury-related or not using ML models trained on diverse datasets.
• Stage 2 predicted body parts injured and injury diagnosis codes from injury-related posts using ML models.

Main Results:

• Stage 1 models (LSTM, GRU) achieved an F1-score of 72% for classifying injury-related posts.
• Stage 2 models (SGD) demonstrated an F1-score of 86% for predicting body parts injured.
• Stage 2 models also achieved a 76% F1-score for injury diagnosis code prediction.

Conclusions:

• The ML framework shows promising accuracy for injury surveillance.
• Social media data analysis via this framework can identify emerging product-related injury trends.
• The proposed system can significantly enhance existing public health surveillance efforts for consumer products.