TL;DR
This paper explores how to automatically identify expressions of medication non-adherence (MNA) and the reasons behind it on Twitter. By analyzing millions of tweets, the authors demonstrate that social media can be a valuable, albeit noisy, source for understanding patient behavior, paving the way for more scalable health research.
What Problem Does This Paper Address?
Medication non-adherence, where patients don't take their prescribed medications, is a significant public health issue, leading to increased illness, mortality, and billions in healthcare costs annually. Understanding why patients don't adhere is crucial for developing effective interventions. However, traditional methods for gathering this information, like clinical records or surveys, are often limited, time-consuming, and may not capture the full picture of patient experiences. The authors propose leveraging the vast amount of user-generated content on social media, specifically Twitter, as a more accessible and potentially richer source of information on MNA and its underlying reasons.
Methods
The researchers employed a two-pronged semi-automatic approach to identify MNA-related tweets within a large corpus.
- Drug Name Matching: This method involved compiling a list of 103 drug names (both generic and brand) related to HIV and diabetes. Tweets containing these drug names were initially retrieved. Heuristic rules were then applied to filter out non-personal content, such as tweets with hyperlinks (often from companies commenting on articles), retweets, replies, and non-English tweets.
- Patterns Matching: This approach used regular expressions (REs) to search for specific phrases indicative of MNA, such as "stopped taking" and "made me" (in any order). These patterns were designed to capture more personal accounts of drug experiences. Importantly, the standard filtering rules for hyperlinks and retweets were not applied here, as the semantic nature of these patterns suggested they were more likely to retrieve personal tweets.
Both methods then relied on manual annotation. Two annotators independently reviewed the retrieved tweets to determine if they expressed MNA. For tweets identified as MNA, a third annotator reviewed the user's timeline (up to ten days before and after the MNA tweet) to identify the reasons for non-adherence.
Datasets
The primary dataset used was the "Pregnancy corpus," consisting of approximately 112,500 timelines (exhaustive sets of tweets from individual users) of women posting during their pregnancy. This corpus was originally collected for a previous epidemiological study. The authors analyzed a subset of 4 million tweets from this corpus. No new benchmark datasets were introduced in this paper.
Key Insights and Ideas
- Twitter as a Rich Source for MNA: The study confirms that users not only express non-adherence on Twitter but also often provide reasons for their behavior. This highlights the potential of social media for large-scale health research.
- Challenges in Automation: While promising, the current semi-automatic methods still require significant manual annotation, indicating that fully automating the extraction of MNA reasons is a complex task.
- Personal vs. Commercial Content: The distinction between personal experiences and company-driven content is crucial. Filtering out promotional tweets (often identified by hyperlinks) is key to isolating genuine user experiences.
- Context is King: Identifying MNA reasons often requires looking beyond the single tweet, examining the user's timeline for preceding or subsequent posts that provide context.
Key Results
The study analyzed 4 million tweets and successfully identified a total of 18 tweets expressing medication non-adherence.
- Drug Name Matching: Retrieved 377 tweets, of which 9 mentioned MNA. Six of these 9 tweets also described the reason for non-adherence within the tweet itself.
- Patterns Matching: Retrieved 27 unique tweets (not overlapping with the drug name matching results), of which 9 mentioned MNA. The precision for this method was 9/27 (33.3%), which the authors note appears more precise than the drug name matching method (9/377, ~2.4%). All 9 MNA tweets identified by this method also mentioned the reasons for non-adherence, with one exception where the reason was in a subsequent post.
- Specificity of Mentions: Of the 9 MNA tweets found via pattern matching, one specified a particular medication, two mentioned a type of medication (e.g., pain medication), and the rest used generalizations (e.g., "pills") or pronouns.
Limitations and Future Work
Limitations: * Small Sample Size: The study analyzed a subset of 4 million tweets, and the number of identified MNA tweets was small (18 in total). * Manual Annotation Burden: The reliance on manual annotation for identifying MNA and its reasons is time-consuming and limits scalability. * Precision of Drug Name Matching: The drug name matching method had a low precision, suggesting a need for more refined filtering or matching techniques. * Ambiguity in Mentions: Many tweets used generalizations or pronouns for medications, making it difficult to identify the specific drug without further context.
Future Work: The authors propose several directions:
- Real-time Listening: Implementing real-time tweet stream monitoring to capture more current MNA expressions.
- Improved Recognition: Developing a more robust "Drug Name Recognizer" and expanding the set of manually curated patterns.
- Full Automation: The ultimate goal is to fully automatize the process of extracting and analyzing MNA reasons to facilitate large-scale epidemiological studies.
Summary updated