Diabetes Decoder

Bridging the medical literacy gap by automatically translating diabetes-related medical jargon into everyday language to empower personal health management

web hero mockup of DiabetesDecoder interface after audio transcription

Role

Product Designer & Researcher

Tools

Python, VS Code, ZeroWidth LLM, OpenAI API

Year

5 weeks - 2024

collaborators

Solo project

Problem

With 25.7 million Americans struggling with limited English proficiency, interpreter shortages cascade into medication errors, extended hospital stays, and 35% worse health outcomes. How might AI-powered tools bridge language gaps and transform patients into confident self-advocates for their diabetes care?

Approach

Built Python-based audio capture system and integrated OpenAI Whisper's multilingual capabilities for transcribing the audio file. Fine-tuned an LLM on diabetes terminology to detect and translate complex medical terminology into accessible plain language. Displayed improved and original transcripts on a multimedia Streamlit interface.

Outcome

Deployed the end-to-end system for validation testing, benchmarking DiabetesDecoder against clinical summaries and generic ChatGPT translations. 85% user preference for DiabetesDecoder's clarity and confidence demonstrates how AI-powered language tools present promising opportunities for inclusive patient empowerment and improved diabetes self-management outcomes.

Workflow Walkthrough

Covering my motivation, building processes, tech stack, and demo

Play icon
Breakdown of Video's Main Components

Peek into Prototyping

This project was broken into 4 major components that are explained below (except for audio recording, which as a temporary local solution was enabled by PvRecorder).
Mockup

Customizing ZeroWidth LLM RAG

Selecting GPT-4 as the base model with its newly raised token count of 128,000 to account for lengthy appointments, I instructed its role as a medical translator and specific knowledge sets to pull from. I embedded the American Association of Diabetes Glossary of Terms as my source, splitting it into 6 knowledge sets to help with search speed and accuracy.

Mockup

OpenAI Whisper: Speech-to-Text

Whisper is a deep learning model trained on over 680,000 hours of multilingual audio data. Given accuracy being the most critical factor in medical settings, I chose this model to create transcriptions from the original audio recording.

Mockup

Multimedia Display Interface

Streamlit is an open-source framework, Python-based library that allows users to create an application with minimal code and supports multimedia capabilities. Because I wanted to create a tool that could be accessible to people of all socioeconomic and cultural groups, I needed to find multimedia ways of presenting information.

Still curious?

Reach out to learn more details about the project—I promise we've got lots more in store:)
Let's connect