Résumé
Background: Cervical cancer is the second most common cancer among women under 45 years of age. To deal with the decrease of smear test coverage in the UK, a Twitter campaign called #SmearForSmear has been launched in 2015 for the European Cervical Cancer Prevention Week. Its aim was to encourage women to take a selfie showing their lipstick going over the edge and post it on Twitter with a raising awareness message promoting cervical cancer screening. The estimated audience was 500 million people. In a previous study (Lenoir et al., J Med Internet Res 19(10):e344, 2017, 10.2196/jmir.8421, http://www.jmir.org/2017/10/e344/), we identified the tweets delivering a raising awareness message promoting cervical cancer screening (sensitizing tweets) and understood the characteristics of Twitter users posting about this campaign.
Objective: The objective of this new study is to investigate the interest of deep learning methods to automatically categorize tweets according to themes and users’ status.
Methods: We conducted a 4-step content analysis of the English tweets tagged #SmearForSmear and posted on Twitter for the 2015 European Cervical Cancer Prevention Week. 18,292 messages were collected using the Twitter Streaming API between the period of January 2017 and November 2017. In order to produce training and test data sets, we annotated the messages according to themes and users’ statuses.
These messages have been analyzed by two independent researchers using a thematic analysis, validated by a strong Cohen kappa coefficient. A total of seven themes were coded for sensitizing tweets and seven for Twitter users’ status. Based on this annotation, we compared by cross validation the predictive performances of traditional classification techniques against more advanced deep learning methods.
Results: Deep learning models were able to predict efficiently the seven themes and seven users’ status. More specifically, the deep learning models performed better than traditional approaches.
Conclusions: Deep learning methods can efficiently predict themes and users’ status. These predictive models could be used as a powerful tool to automatically analyze social data such as twitter streams for medical perspectives. This study also demonstrates that the success of a public health campaign using a social media platform depends on its ability to get its targets involved. It also suggests the need to use social marketing based on efficient predictive approaches to help its dissemination. The clinical impact of this Twitter campaign to increase cervical cancer screening is yet to be evaluated.