← Research notebook

Reading a post for reported drug reactions

A step-by-step sketch of the general workflow my undergraduate research studies: finding drug and reaction mentions in informal text, pairing them, and checking each pair against what the writer actually said.

Illustration only: the posts and drug names are invented, and no real data, models, or results are used.

Step through an example

Example 1: a clear first-person report

example post

I started taking Drug A last month and noticed headaches afterward.

  • Drug A → headache Supported by the text The writer reports taking the drug and the symptom, in that order (“afterward”).
Example 2: two drugs, one reaction that went away

example post

Switched from Drug B to Drug C. The nausea I had on Drug B is gone, but now I can’t sleep.

  • Drug B → nausea Supported by the text “The nausea I had on Drug B” ties the symptom to that drug.
  • Drug C → insomnia Supported by the text Sleep trouble is reported “now”, after the switch to Drug C.
  • Drug C → nausea Contradicted The text says the nausea is gone on Drug C. Pairing every drug with every symptom would get this wrong.
Example 3: someone else’s symptom and a hedged one

example post

My sister gets dizzy on Drug D. I’ve been on it a week and feel fine, just a bit tired, probably from finals though.

  • Drug D → dizziness Reported about another person The writer describes their sister. Whether that counts depends on the annotation guidelines, so it is flagged rather than accepted.
  • Drug D → fatigue Uncertain: abstain The writer offers another cause (“probably from finals”). A careful system can decline to assert a pair.

  1. 1. Original text

    The input is informal writing. People use brand names, slang, typos, and hedges, and they describe other people’s experiences as well as their own.

  2. 2. Drug mentions

    The first pass marks spans that name a medication. In the real setting this means handling spelling variants and mapping names to a standard vocabulary.

  3. 3. Reaction mentions

    The next pass marks spans that describe a symptom or side effect. Casual phrases are mapped to a standard term, shown in small type, so “I can’t sleep” and “insomnia” can be compared.

  4. 4. Candidate pairs

    Drug and reaction mentions in the same post are joined into candidate pairs. A candidate is only a hypothesis to check; at this point nothing has been decided.

  5. 5. Review against the text

    Each candidate is checked against the words that support or contradict it. Pairs can be supported, contradicted, flagged for a human decision, or left out when the evidence is too weak. The research plan compares direct prompting, retrieval-augmented generation, and extraction with verification against human annotations.

  6. 6. A report is not proof of causation

    Even a well-supported pair only says that someone reported a reaction while taking a drug. It does not show the drug caused it: the timing could be coincidence, other medications or conditions could be involved, and online posts are not a representative sample. Extracted pairs are signals for further study, not medical findings, and nothing on this page is medical advice.

Key

Read about the research →