Clinical trial search in plain language.
A search platform that lets patients describe their condition in their own words and matches them to suitable clinical trials, with no medical terminology required. We reworked the prompt design so output fits a structured search display.
- Client
- Duke-NUS
- Sector
- Software & AI
- Year
- 2024
- Our role
- Design, LLM integration and web app
- Duration
- 2 months
Patients do not speak in trial codes.
Finding a suitable clinical trial usually means knowing the medical terms trials are listed under. Our client at Duke-NUS wanted patients to describe their condition and preferences in their own words and still find the trials that could help them.
The language model behind the search ran on the ChatGPT API, but the original prompts returned free text that was hard to place into a fixed search display.
Prompt, interface, deploy.
- 01
UI and UX
The patient describes their condition in their own words and the model turns it into a short profile they can check and correct. It then searches the trial documents, checks the patient against each trial's criteria and ranks the results, showing why each one may suit them: met, to confirm or not met, with a link to the document it came from. Each trial opens with a plain language summary, what taking part involves and questions to ask the study team, on desktop, tablet and phone.
- 02
Prompt design
We studied the original prompts and reworked them so the output arrives in a fixed structure, tuned for both quality and cost per query.
- 03
Web app
Front end and back end built around the model, with tools for inspecting and debugging its output.
- 04
Deployment
The platform runs on AWS.
Delivered in two months.
The platform and its debugging tools were designed and built in two months. It lowers the barrier for patients looking for trials and cuts the time medical staff spend matching patients to suitable studies.