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

Safety is a critical aspect of consumer products. However, there are millions of product related injuries reported each year. Traditional injury surveillance efforts led by public health agencies involve product related injury data collection from hospitals and subsequent injury causation analysis. This approach often requires long processing time and leads to delays in identifying emergent consumer product-related injury patterns and preventive intervention steps such as product recalls, which causes continued product related injuries.

We propose a machine learning (ML) based framework for improving injury surveillance by extracting crucial product injury related details from real-time social media posts to quickly identify emerging trends of product injuries and potentially facilitate timely interventions. We evaluated the efficacy of the proposed framework by analyzing injuries related to skateboard from the Redditt platform. The framework has two stages. In stage1, the social media posts scrapped based on product-related keywords were classified whether they were injury related or not using ML models trained on non-injury related data of Amazon product reviews and injury-related data obtained from the National Electronic Injury Surveillance System (NEISS) database. In stage2, the posts identified as injury related were further analyzed by another ML model trained on NEISS dataset to predict the body-part injured and the injury diagnosis code based on the content of Redditt post.

In stage1 for classifying whether social media posts are injury related or not the deep learning models LSTM and GRU yielded an F-1 score of 72 %. In stage2, for the posts that were classified as injury related, the SGD model yielded an F-1 score of 86 % for predicting the body-part-injured and 76 % for injury diagnosis-code.

The results of the study indicate that the proposed machine learning framework yielded decent accuracy levels for injury surveillance purposes. Therefore, the framework can be used for analyzing social media data for identifying emerging trends in product-related injuries and can bolster the existing injury surveillance efforts.