Building AI-Powered Health Insights
The Challenge: AI Integration to PHR Solution
Fasten Health (FH) emerged as the standout entry in our 2024 AI idea contest, which sought innovative AI integrations for existing products.
FH offers a personal health record (PHR) solution to democratize access to medical information. Their platform addresses a common problem: patients' health records are often scattered across various healthcare organizations, leaving individuals needing a complete picture of their own health history. FH's mission is to create a "unified, accessible hub tailored for individuals and families" while adhering to strict data privacy standards (HIPAA compliance, end-to-end encryption, and on-device processing).
The proposed AI feature would enhance FH's offering by integrating an AI-powered assistant that transforms raw health data into understandable, actionable information, empowering users to take a more active role in managing their health.
SERVICES
AI / ML
Product Dsicovery
Product Design
UX/UI
ENGAGEMENT
Ongoning
PLATFORM
Figma
GitHub
The Process: Product Discovery and Technical Feasibility
Traditional product development typically begins with product understanding, including market position, capabilities, and competitor analysis. This is followed by idea generation and opportunity identification. Only after these stages does the evaluation of potential technologies or solutions, such as AI, come into play. However, our engagement with Fasten Health deviated from this norm due to several unique factors.
We approached this project with existing knowledge of Fasten Health's product, a clear requirement for AI integration, and a solid understanding of our AI capabilities and areas for exploration. This unusual starting point allowed us to pursue a more streamlined development process. Instead of following the typical sequential approach, we conducted product discovery in tandem with assessing technical feasibility.
Product Discovery
Product discovery is not a one-size-fits-all process. The approach varies significantly depending on whether we're starting with a bare concept, a functional product, or a well-researched idea. In essence, product discovery is as diverse as the clients and projects we encounter.
For Fasten Health, we found ourselves in an interesting position. While we were aware of their existing product vision, including their competitors and market understanding of competitors, the new AI feature was uncharted territory. This became the focal point of our discovery sessions.
Our product discovery process unfolded over four targeted workshop sessions, each designed to build upon the last and bring us closer to a clear, actionable vision for the product as a whole. Throughout this discovery process, we employed a variety of tools and techniques, each chosen to extract specific insights and drive us toward our goal of a well-defined, user-centric AI feature for Fasten Health.
Architecture definition:
We proposed a modular architecture capable of search and Q&A functionalities to integrate AI into Fasten. This architecture is known as RAG (Retrieval Augmented Generation) in the AI field. This architecture includes:
- Text-to-embedding model: Converts user queries and medical documents into embeddings for efficient search and retrieval.
- Search module: A combination of vector and full-text search is used to identify the most relevant documents based on user queries.
- Text generation model: Leverages retrieved documents to generate accurate, contextually relevant answers to user questions.
Two deployment options were outlined: an entirely local setup or a hybrid approach leveraging cloud services for model inference. The architecture is designed to be flexible, allowing local deployment to maintain data privacy while offering the option to offload resource-intensive tasks to the cloud.
This translates into Fasten having a 'progressive privacy level', allowing users to use a third party (better performing, less private), run AI locally (more private, less performant), or opt out of AI features completely.
Technical Feasibility
There was an existing idea of what the feature had to be like: an AI assistant that run locally in any user's device to protect data privacy and sensitive information. To assess technical feasibility and move along those boundaries, we devised the following stages:
Model research:
This stage involved an in-depth investigation of the top Large Language Models (LLMs) tailored to the medical field. The research focused on text-based and multimodal models, evaluating their performance across medical datasets included in the Open Medical LLM Leaderboard:
- MedQA
- PubMedQA
- MMLU
The models were then benchmarked based on the leaderboard's average scores as well as several other parameters; for example, they shouldn't go over 8B and should have a b16 in terms of precision. This was key for the models to perform locally as the team emphasized balancing model precision and resource requirements for efficient local execution on users' systems.
Future work and roadmap:
The team gathered a list of future work to continue after the Discovery process, which involves building a robust evaluation framework to validate model performance across key metrics, enhancing generative answers through new models, and optimizing system responses by refining hardware configurations and prompt engineering.
We developed a roadmap prioritizing system evaluation and improvements before making external changes. This approach ensures we won't integrate or update the user interface of a system that isn't yet performing optimally. It also allows us to confirm that our local system is both accurate and fast enough for practical use.
The Design Track
The design prioritizes consistency with Fasten Health's existing aesthetic. It includes a toggle to turn off the AI features with a single click, efficiently addressing data privacy concerns. It also provides users with suggestions of possible prompts to assist them in using the virtual assistant and ensure they take full advantage of its potential.
Other features include the option to adjust tone, providing users with a better understanding of complex medical jargon, and exam results display and its relation to other health aspects.
Results: Fasten Answers, an AI assistant integration
The process for Fasten Answers demonstrated the critical role of thorough research and validation in AI implementation, whether enhancing a feature or creating a new product. Our team researched technical feasibility, product-market fit, and consistent design. The process also enabled product owners to refine key elements like competitor analysis, personas, and user journeys. Ultimately, we delivered a comprehensive roadmap detailing each step required to deploy the solution, paving the way for its successful implementation.