Personalized Medical Records

Inspiration

Personal Medical records are not easily consumable by users - we want to unlock the capability to deeply understand your own health data longitudinally, and also consume the data to generate recommendations, activity and plans

What it does

The app consumes FHIR medical record and answers free form questions on the medical record. It can answer questions summarizing your conditions, diving into specific conditions, diagnostic plan, Medications and also make recommendations based on the recent observations.

How we built it

It's built using langchain, openi and prompts.

The prompt is generic - i.e. it can work against any resource by giving few shot examples, this makes the implementation very flexible to the dozens of resource types in the FRIC spec

Challenges we ran into

The main challenge is that FHIR is a big record and exceeds token length, so it requires some steps to get to the data to extract the right objects and then query them.

Accomplishments that we're proud of

The main accomplishment is identified an underserved area, to the best of my knowledge hackers haven't looked into applying GPT to personal records, this will open up more opportunities as this is developed further.

Addtionally, It's an end to end functional app and is already able to make personalized suggestions that are useful

What we learned

Primarily, the difference between GPT-4 and GPT-3. I spent some time working with GPT-3 earlier, this time could have been better spent working on GPT-4 directly since it's much more powerful.

I spent a lot of time trying advanced patterns such as chain of thought etc, but it turned out that 2 prompts one after the other just worked - all I had to do was parametrize the first prompt and give some few shot examples.

What's next for Personalized Medical Records

This work requires a Safety layer, since it can be misused and misinterpreted, so it should not be used without those layers.

Built With

• javascript
• langchain
• openi