The problem ARISE solves

Individuals spread across a number of healthcare systems face the issue of aggregating their data records into a single location while also keeping track of the variety of file formats, doctor reccomendations, second consultations and obtaining health suggestions to their real time habits. After consulting with healthcare professionals, the top three issues with documentation existed of uploaded PDFs, lack of searchability due to incomplete records and fragmented records across multiple platforms.

Through the development of ARISE, we targeted how we could demonstrate aggregating this data collectively while powering further health suggestions past basic information through generative AI. Furthermore, we wanted to actively funnel this data into a custom user portal for easy access when documentation of medical history are required. We used Supabase for easy registration and user authentication, which then linked to a Postgres SQL database sorted by user ID. PDFs and reports are also stored on Supabase's Cloud Storage, for each access of individual reports. Though we weren't able to enable parsing for all file formats, it was a testament in converting data into a single file system for EHR data. We used OpenAI's GAI model to provide cross-references in data and general insights that may spark red flags, and also created a suggestive nutrition schedule. For parsing the PDFs, we used Open AI's GAI api, Python, and FastAPI to return, host and endpoint to access the assigned data.

Furthermore, analyzing numerous page long after-visit summaries is a tedious process, which we aimed to make engaging. With studies demonstrating only 50% information retention, we wanted to enable customization in listening to the most important parts of your own health records. Thus, we decided to use Open AI's GAI api to generate an engaging song by the singer of your choosing. Unfortunately, due to time constraints and access to data, we only developed the model for Drake.

Challenges we ran into

We faced challenges with making the OpenAI Model we chose to conform to our strict JSON format. Numerous times, the model would output results that were slightly off from the format we expected. So, we researched online through multiple articles and got consistent results by rolling back the API version we were using and utilizing the “functions” feature.

In addition, we had trouble making our AI voice match Taylor Swift’s song style. We realized that voice-morph AI solutions focus on rap-genre music. So, we focused on aligning the voice with Drake’s lyric style, to which we succeeded greatly. Finally, we encountered challenges with Python and FastAPI since it was our first time developing a web backend. However, reading the documentation carefully solved most of our problems. Overall, the project was a fun learning experience where we learned a lot about the current state of AI and full-stack development.

Additionally issues we faced was the lack of access to medical records and real medical data enabled by companies like Epic systems and other major healthcare infrastructure providers. As a result, we wanted to adapt for one of the data types that was commonly found to be an issue. We decided to target PDF post-appointment reports, and parsing this data into our database. However, we weren't able to target all data forms that would be realistic of the actual industry due to lack of accessibility. Another issue was primarily in the ideation phase as majority of our problems stemmed from this lack of data.

On the software development end, we found issues integrating dynamic UI as the API would sometimes not generate certain fields depending on the type of file that was fed to it. We solved this by conditionally displaying data based on the output of the Open AI GAI api. Additionally, the back and forth between processing the mp4 speech was something that not completely resolve and took an unideal amount of time to retrieve.