Inspiration
The theme was accessibility in navigating the medical system, so I started by looking at what actually stops disabled patients from getting care. One problem kept coming up: inaccessible medical equipment. Fixed-height exam tables, standard scales, no patient lifts. Wheelchair users are often examined sitting in their chair, or not weighed at all, which affects diagnosis and medication dosing.
Then I tried to find out which Atlanta hospitals have accessible equipment. I checked the official websites of Emory, Piedmont and Grady hospitals. None of them published it. Patients only find out when they arrive.
Reviews don't fill the gap either. A one-star review might be about parking while the lift worked perfectly, and a five-star review might come from someone who couldn't get onto the exam table. One star rating can't tell a wheelchair user whether the equipment they need works.
What it does
HospitalRating shows, for every place and every piece of accessibility equipment, what the hospital says next to what patients say, and warns when the two don't match.
- Choose your needs: adjustable exam table, wheelchair scale, patient lift, accessible restroom, step-free entrance, staff help, accessible mammography.
- Browse 23 real places in metro Atlanta and Gwinnett: 8 hospitals, 7 urgent care centers and 8 specialist clinics (ENT, cardiology, orthopedics, breast imaging).
- See match scores like "Patients confirm 3 of 4 of your needs."
- Write a review in your own words, typed or spoken. The app reads it sentence by sentence and updates the results live. Stars are never used.
- Call the clinic or get directions in Google Maps or Apple Maps with one tap.
How I built it
- Plain HTML, CSS and JavaScript. No frameworks, no server, no accounts, so it runs anywhere, even offline, and is hosted on GitHub Pages.
- Machine learning written from scratch: two Naive Bayes text classifiers in JavaScript. Each review is split into sentences. The first classifier decides which equipment the sentence is about (or "other," like parking, which is ignored). The second decides whether it worked or failed. If confidence is below 60%, the app asks the user to confirm instead of guessing, and users can correct mistakes with one tap.
- Training data: 162 labelled sentences written in the casual, run-on, sometimes misspelled way people actually write reviews.
- Evaluation: a Node.js script holds out 20% for testing. Item accuracy is 93.8%, and 96.6% when the model is confident. Worked/failed accuracy is 81.3%.
- Voice input uses the Web Speech API, with clear error messages and a typing fallback.
- Real data: hospital names, addresses and phone numbers were checked against official websites. Places I couldn't check are tagged "unverified."
- I used Claude Code, an AI coding assistant, to help plan and write the code.
Challenges I ran into
- The data doesn't exist. I couldn't find any public source for hospital accessibility equipment. That became the core of the project instead of a blocker: every hospital honestly shows "Not published," and patients fill the gap.
- Staying honest with real hospitals. Made-up negative reviews about real hospitals would be false claims. All demo reviews are clearly labelled, and I never invented an equipment claim.
- Verifying data: some hospital websites blocked my checks, so those places are marked unverified instead of guessed.
- Voice input only works in some browsers and only over a secure connection, which took debugging.
- Worked vs. failed is harder than it looks. Sentences like "the table was supposed to lower but nobody knew how" mix both, which is why that classifier is less accurate than the item classifier.
- Changing direction: I explored several other ideas before landing on this one, with limited time and working solo.
Accomplishments I'm proud of
- Machine learning I built and can fully explain, with real accuracy numbers, including where it gets confused.
- Turning a 1-star review into useful equipment information ("lift: worked") instead of a penalty.
- An app that is accessible itself: large text with size controls, keyboard navigation, screen reader support, and icons plus words instead of color alone.
- Using real hospitals while staying honest about what's demo data.
What I learned
- Accessibility problems are often information problems: the equipment might exist, but nobody tells the patient.
- A single rating hides the details that matter most to disabled users.
- A small, well-chosen ML model you can measure is more useful, and easier to trust, than a big one you can't explain.
What's next
- Hospitals publishing their own equipment data. The 2024 federal rule on accessible medical equipment gives them a reason to.
- Confirming reviews come from a real visit, to stop fake reviews.
- More training data from real patients, more languages, and coverage across Georgia.
Log in or sign up for Devpost to join the conversation.