Automatically Extracting Patient Opinions from Twitter

To investigate the potential for identifying patient opinions from general social media data, without requiring mention of official accounts; and to investigate the differences in polarity and topic of opinion that this produces.

We used supervised and unsupervised machine learning methods for the extraction of patient opinion data from social media. We investigated the feasibility of extracting insightful data from social media that can provide indicators of patient care performance in healthcare services.

Previous studies had relied on direct mentions (i.e using @hospital_handle), which leads to a large portion of data being missed. We collected tweets using keyword search, and compared tweets that featured direct mentions to those that didn’t. We found that users were more likely to include direct mentions in tweets with a positive sentiment, such as thanking staff for taking care of them.

Broadening out the stream of social media data using a keyword-driven approach yields a bigger pool of patient opinions, which is less biased towards the positive – it can therefore help hospitals find more reports of negative experience in a broader range of areas. More care must be taken filtering data for relevance, but automatic filtering methods are suitable; although some manual post-filtering or periodic retraining of automatic tools is likely to be required.