Dental Sentiment Analysis
NLP pipeline that turns patient reviews into actionable clinic insights
An NLP application that analyses dental patient reviews and classifies sentiment to help a clinic surface patterns in customer feedback — satisfaction trends over time, recurring complaints, and service quality signals that were previously invisible.
What needed to change
A dental clinic had years of patient reviews scattered across Google, health directories, and its own feedback forms. Reading them individually was slow, and no one had time to look for patterns across hundreds of reviews. The clinic manager suspected there were recurring complaints about wait times and billing clarity, but couldn't confirm without a systematic analysis.
What made it hard
- Patient reviews are short, informal, and often sarcastic — naive sentiment analysis fails badly on them.
- The clinic wanted a live demo, not a one-off report — so the pipeline had to be deployable.
- No cloud budget — the solution had to run on free infrastructure.
- The clinic had no in-house data team — the output had to be understandable to a non-technical manager.
How we built it
We built a full NLP pipeline: text cleaning, tokenisation, stop-word removal, and TF-IDF vectorisation, followed by a sentiment classifier. We compared three approaches — VADER (rule-based), a Naive Bayes model, and a logistic regression model — using a small hand-labelled set of clinic reviews as ground truth. The logistic regression model with TF-IDF features performed best on the informal review corpus. We deployed the whole thing as a Streamlit app on Hugging Face Spaces, which gives the clinic a public URL they can share with staff — no cloud account needed.
What it does
- Automated sentiment scoring (positive, neutral, negative)
- Trend detection over time (by month, by quarter)
- Common complaint identification via keyword extraction
- Service-quality insights dashboard
- Public live demo (Hugging Face Spaces)
- Reproducible notebook for retraining on new reviews
What changed
The clinic manager confirmed two of the three top complaints were the ones they'd suspected (wait times, billing clarity) and were surprised by the third (a specific procedure that generated disproportionately negative sentiment). They now review the dashboard monthly and have adjusted their front-desk flow to reduce wait times. The Hugging Face demo has also become a portfolio piece — it's the project most referenced by prospective clients who want to see our NLP work in action.
What I'd do differently
The hardest part was honest labelling. Dental reviews are short, and a two-sentence review can be positive about the dentist but negative about the parking — meaning a single 'sentiment label' is often wrong at the review level even when it's right on average. In a future version, I would switch to aspect-based sentiment analysis (separate scores for treatment quality, staff friendliness, wait times, pricing) — that's what the clinic actually cares about. I'd also add a feedback loop so the clinic can correct misclassified reviews and improve the model over time.
Need something like this?
Whether it's a web platform, an AI pipeline, or a full solar installation — we build solutions tailored to your exact needs.