← Back to WorkData Science

Drug Reviews NLP

I studied more than 215,000 historical patient reviews to understand which medications people rated highly, where patients repeatedly described poor experiences, and how a comparison tool could make those patterns easier to explore.

RoleData Analyst
StackPython, SQL, NLTK, Streamlit
Timeline2025
215K+
Patient Reviews Studied
3,436
Medications Covered
885
Health Conditions
$27M+
Estimated Revenue Risk

Overview

The Question

People taking medications for the same condition can report very different experiences. I wanted to know whether patient reviews could reveal where those differences were largest and make better-rated alternatives easier to find.

What I Did

I organized reviews collected from 2008 to 2017, compared medications used for the same condition, and looked for repeated concerns in the written feedback. Then I built an interactive tool that lets someone explore those patterns without reading thousands of reviews.

Making Thousands of Reviews Easier to Compare

Instead of presenting a wall of analysis, the tool starts with one condition and shows how patient ratings differ across the medications used to treat it.

Historical patient review data

Drug Alternative Finder

Condition: Depression
11,640patient reviews
50medications compared
4.7 pointsbetween the highest and lowest average ratings
Highest average patient rating9.4 / 10
Lowest average patient rating4.7 / 10
Why this matters

The large gap shows where patients reported very different experiences with medications used for the same condition. It creates a useful starting point for questions, not a medical recommendation.

Based on historical patient reviews from 2008–2017. For informational analysis only; medication decisions should be made with a healthcare professional.

Key Findings

Ratings Varied Widely Within the Same Condition

For several common conditions, the highest- and lowest-rated medications were separated by five points or more on a ten-point scale.

Eight Conditions Had Consistently Low Ratings

Across these conditions, the average patient rating stayed below 6 out of 10, suggesting repeated dissatisfaction rather than one poorly rated medication.

Birth Control Generated the Most Feedback

It represented 18% of all reviews but still received below-average ratings, making it one of the clearest areas for deeper research and product improvement.

Written Feedback Added Context to the Score

Some people selected a high rating while still describing concerns in their written review. Reading both signals helped surface dissatisfaction that a score alone could miss.

How I Worked Through It

Start
Patient reviews
215K experiences
Organize
Group the data
By condition and medication
Compare
Ratings and words
Find repeated patterns
Understand
Find the gaps
See where experiences differ
Share
Comparison tool
Make findings explorable

The Work Behind the Findings

1Reading More Than the Rating
What I didI compared each person's written feedback with the score they selected. This made it possible to flag reviews where the words sounded much more negative than the rating suggested.
Why it matteredIt uncovered concerns that would have been missed if the analysis only looked at the numeric rating.
2Finding Repeated Patient Concerns
What I didI grouped commonly repeated words and phrases in negative reviews to see which concerns appeared again and again across medications and conditions.
Why it matteredIt turned thousands of individual comments into a clearer picture of the issues patients mentioned most often.
3Connecting Patient Experience to Business Risk
What I didI combined review patterns with outside research about medication discontinuation to estimate how poor patient experiences could affect continued use and revenue.
Why it matteredThe model estimated more than $27M in potential revenue risk, while making the assumptions and need for further validation clear.

Explore the Full Tool

The interactive version lets visitors choose a condition, compare patient ratings across medications, review commonly mentioned concerns, and read examples of positive and critical feedback.

Try the Live App ↗
Choose a conditionCompare patient ratingsReview common concernsRead sample feedback

Impact

Revenue Risk
$27M+
Reviews Analyzed
215K
Conditions Flagged
8