A social-media methods paper separated medication misuse language from personal-use reports and unrelated mentions.
The annotated dataset
Researchers created a corpus of 16,443 tweets mentioning at least twenty medicines, including benzodiazepines. An iterative process refined categories covering abuse or misuse, personal consumption, mentions, and unrelated material. Final interannotator agreement was 0.86 by Cohen’s kappa. The best reported classifier achieved about 73% accuracy.
High agreement shows consistency between annotators under the adopted definitions. It does not establish that every statement was true or that the person posting actually consumed the medicine.
Limits for an individual investigation
The study concerned Twitter text and selected medicine names. Language, humor, quotation, ambiguity, and the difference between an account and a person constrain exposure inference. Aggregate classifier accuracy also does not establish the reliability of one classification.
For evidentiary use, distinguish the original words, the annotation decision, the algorithm output, and independent evidence of conduct. A post referring to alprazolam may be relevant context, but it is not a biological test or a dose history. The paper contributes reproducible classification methods for surveillance rather than a method of proving intoxication, misuse, or simultaneous ingestion in an individual case.
Source paper
Original analysis of the supplied journal article, identified as PMC7066507.
Leave a Reply