A word can become a drug reference without changing its spelling. It can mean one substance in one online community, another in a different city and nothing drug-related at all in an ordinary conversation. Emoji, abbreviations and deliberate misspellings make the problem harder. By the time a static glossary is updated, the language that matters may already have moved.[1]
A new perspective in npj Digital Public Health proposes an ontology-enhanced large-language-model pipeline for this surveillance gap. The authors do not present a finished detector. They describe how retrieval, knowledge graphs and expert review could prevent a language model from improvising public-health categories while still allowing the vocabulary to evolve.[1]
Why the surveillance problem is consequential
| Measure | Value | Context |
|---|---|---|
| Long-run mortality context | US drug-involved death rate about 250% higher than 20 years earlier | The paper contrasts this long rise with a 24% decline from 2023 to 2024. |
| Nonfatal burden | 8.4 million drug-related emergency visits in 2024 | Reported from the US Drug Abuse Warning Network in the perspective. |
| Language source | Posts, slang, codes and emoji | Signals appear before they are standardised in clinical or administrative records. |
| Paper type | Conceptual perspective | No new dataset was analysed and no source code was released. |
An ontology gives the model something firmer than resemblance
A glossary lists terms. An ontology records relationships: a term may refer to a substance, a route of use, a dosage pattern, a community or a time period. A knowledge graph can express those relationships as subject–predicate–object triples. Retrieval then supplies the model with the approved meanings that are relevant to a post instead of asking it to rely on whatever associations happened to be learned during pretraining.[1]
A controlled route for adding new language
| Stage | System action | Human responsibility |
|---|---|---|
| Retrieve | Find ontology entries and graph triples matching the post | Define trusted sources and acceptable semantic relationships. |
| Interpret | Use the retrieved context to classify or explain the language | Check whether the model has preserved uncertainty and context. |
| Flag unmatched terms | Treat unexplained slang as a candidate referent | Investigate frequency, community, geography and co-occurring evidence. |
| Adjudicate and update | Add a term only after review with date and provenance | Prevent one model guess from becoming an official category. |
“Tea” is not evidence until context makes it one
The perspective uses ambiguous language to show why keyword surveillance fails. A common word such as “tea” can be literal, metaphorical or community-specific. The system would need surrounding terms, user context, time and links to known concepts before treating it as a drug reference. High sensitivity without that restraint would flood analysts with ordinary conversation.
An LLM by itself is flexible but unstable: it can recognise novel phrasing, yet it may invent a connection or import an outdated meaning. An ontology by itself is controlled but slow: it cannot contain a term that experts have not observed. The proposed architecture assigns each component a different job—generation for interpretation, retrieval for grounding, graph structure for consistency and people for authorisation.
Surveillance can harm the communities it observes
A system that flags coded language may help identify an emerging substance or adulterant. It can also amplify stigma, draw enforcement attention to marginalised communities or mistake in-group speech for evidence of criminal behaviour. Provenance, access controls, false-positive review and clear limits on downstream use are therefore part of the surveillance design, not an optional ethics appendix.
The useful idea is that vocabulary growth should become an accountable process. A public-health system should be able to say not only that a phrase was flagged, but which approved relationships supported the interpretation, what remained unknown, who reviewed the candidate and when the meaning was added. Drug slang moves quickly; institutional memory should not move invisibly.
Sources & further reading
- Closing surveillance detection gaps for emerging digital drug discourse with ontology-enhanced large language models npj Digital Public Health · 16 Oct 2026 · Accessed 16 Oct 2026