Automating Detection of Drug-Related Harms on Social Media: Machine Learning Framework
Bibliographic record
Abstract
BACKGROUND: A hallmark of unregulated drug markets is their unpredictability and constant evolution with newly introduced substances. People who use drugs and the public health workforce are often unaware of the appearance of new drugs on the unregulated market and their type, safe dosage, and potential adverse effects. This increases risks to people who use drugs, including the risk of unknown consumption and unintentional drug poisoning. Early warning systems (EWSs) can help monitor the landscape of emerging drugs in a given community by collecting and tracking up-to-date information and determining trends. However, there are currently few ways to systematically monitor the appearance and harms of new drugs on the unregulated market in Canada. OBJECTIVE: The goal of this work is to examine how artificial intelligence can assist in identifying patterns of drug-related risks and harms, by monitoring the social media activity of public health and law enforcement groups. This information is beneficial in the form of an EWS as it can be used to identify new and emerging drug trends in various communities. METHODS: To collect data for this study, 145 relevant Twitter accounts throughout Quebec (n=33), Ontario (n=78), and British Columbia (n=34) were manually identified. Tweets posted between August 23 and December 21, 2021, were collected via the application programming interface developed by Twitter for a total of 40,393 tweets. Next, subject matter experts (1) developed keyword filters that reduced the data set to 3746 tweets and (2) manually identified relevant tweets for monitoring and early warning efforts for a total of 464 tweets. Using this information, a zero-shot classifier was applied to tweets from step 1 with a set of keep (drug arrest, drug discovery, and drug report) and not-keep (drug addiction support, public safety report, and others) labels to see how accurately it could extract the tweets identified in step 2. RESULTS: When looking at the accuracy in identifying relevant posts, the system extracted a total of 584 tweets and had an overlap of 392 out of 477 (specificity of ~84.5%) with the subject matter experts. Conversely, the system identified a total of 3162 irrelevant tweets and had an overlap of 3090 (sensitivity of ~94.1%) with the subject matter experts. CONCLUSIONS: This study demonstrates the benefits of using artificial intelligence to assist in finding relevant tweets for an EWS. The results showed that it can be quite accurate in filtering out irrelevant information, which greatly reduces the amount of manual work required. Although the accuracy in retaining relevant information was observed to be lower, an analysis showed that the label definitions can impact the results significantly and would therefore be suitable for future work to refine. Nonetheless, the performance is promising and demonstrates the usefulness of artificial intelligence in this domain.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.010 | 0.006 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.001 | 0.000 |
| Research integrity | 0.001 | 0.008 |
| Insufficient payload (model declined to judge) | 0.003 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one teacher head, not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".