Back to Projects

Pharma NLP

What Patients Say That Ratings Miss

NLP on 215K Drugs.com reviews

  • NLP
  • Python & VADER
  • Pharma Insights

See the work

GitHubKaggle notebook
01

The Problem

Two drugs can both average 7/10 while one keeps giving people headaches or weight gain, and you'd never see it from the rating alone. I wanted to find the side effects patients actually complain about, see which go with bad ratings or with people stopping the drug, and spot drugs that look fine on ratings but hide a problem in the text.

02

Strategy & Solution

  • Cleaned 215,063 reviews down to 127,937 unique ones, removing junk conditions and brand/generic duplicates.
  • Built an adjusted VADER that ignores illness words like "pain" and "depression", which were making pain patients look unhappy just for naming their condition.
  • Detected 15 side effects with negation handling, then compared bad-rating and stop/switch rates with and without each one.
03

Impact

  • Found 35 hidden signals: drugs rated at or above their condition's median with a side effect at 1.5x+ the condition average.
  • Heart racing, skin problems and mood swings raised the odds of a bad rating by 1.6x to 1.8x.
  • Hair loss, mood swings and heart racing were the side effects most tied to patients stopping or switching.

Interested in working together?

Get in touch