Have you ever wondered how they come up with names for prescription drugs? They seldom describe the function of the medicine or the chemical structure. They are seemingly pulled out of someone's... head.
In this project. I create a Markov chain model that generates fictitious medicine names based on real drug brand names gathered from the NDC (National Drug Code) database. Another model generates descriptions from text data gathered from Wikipedia.
After parsing drug brand names from NDC database, descriptions were gathered by through the Wikipedia API. Many requests resulted in errors because the API takes percise names; many names were duplicates. For all successful requests, a sentence starting with "used" was extracted from the first paragraph of the article. Almost all articles have this sentence that describes the use of the drug.
I wrote a Medium article about this project and Markov Chains more broadly. You can find it here. The article goes into greater detail on how the random walk of the Markov chain works (also the code is more up to date there). It's a great place to start if you want to follow along with the project.
I wrote a webapp that uses this API in React. You can check it out here. The source code is also on GitHub.
Parsed data (both descriptions and drug names) are available in the data directory of the repo. For drug brand names, raw data was collected from the NDC Database. Description data was parsed from Wikipedia articles accessed with the Wikipedia API. Raw data (may be a bit out of date) is in data/product.txt. The parsed drug names are in data/names.txt; the descriptions are in data/descriptions.txt.
- Node.js the runtime of the API.
- Python for getting data from Wikipedia, building the chain, and prototyping the generator.
- Java for parsing NDC database (because of streams API).
- Serverless for boostrapping the API
- AWS Lambda for hosting.