HealthECHO

Inspiration

What inspired us is the personal stories surrounding misdiagnoses or dismissive initial consultations experienced by people close to us. In many cases, critical conditions were overlooked because the right questions weren’t asked, or certain aspects weren't scrutinised thoroughly. We understand the immense workload on doctors who see hundreds of patients and can't delve into every single detail. This is where we saw a need for HealthEcho - to empower doctors to provide each patient with the personalised attention and analysis they deserve.

What it does

HealthEcho is an assistant that gives non-intrusive suggestions of potentially overlooked diagnostic angles to doctors following an initial consultation. We focus on suggestions after doctor-led consultation rather than initial screening because we do not want to create an attention bias for the clinician. We leverage the powerful natural language understanding capabilities of Claude-2 to analyse the doctor-patient interactions, extract relevant medical knowledge, and engage doctors in insightful dialogues upon request. To enhance the medical knowledge of Claude-2, we leverage trust-worthy medical resources such as National Institute for Health and Care Excellence (NICE) guidelines, GP Notebook and textbooks to enrich the context. We combine traditional word-based search with vector search to ensure a diversity of relevant information is retrieved.

How we built it

Our approach was top-down, starting with a grand vision, working down to the MVP. We rapidly iterated to leverage the best frameworks for this challenge, bridging the front-end interface with our LLM backend to craft a tangible MVP showcasing the value and impact of HealthEcho. Our system architecture comprises two chains in the LLM backend. The first, an extraction chain, transforms raw transcripts from the consultation into powerful encodings to fetch relevant domain knowledge efficiently from trustworthy medical resources. The second chain leverages this domain knowledge to scrutinise the clinical interaction, highlighting areas needing further exploration that are then suggested to the doctor in charge. Domain knowledge is sourced through three data collection methods:

Certified Textbook Vector Search via MongoDB and Biomedicine BERT.
Global Awareness with Brave WebSearch.
Medwise.ai for access to clinical guidelines.

Challenges we ran into

Navigating through LangChain prompt customization and multi-retrieval issues were significant hurdles considering we wanted to integrate three different data sources. Deciding on relevant data sources and dealing with scope revisions were other significant challenges that we faced. Finally, balancing the complexities of linking the user interface with our LLM backend was also crucial to ensuring a user-friendly and reliable MVP. Constructing a searchable database of indexed content from Medical Textbooks required thoughtful parsing of textbook data and innovative hierarchical database structuring.

Accomplishments that we're proud of

Our proudest achievement is the robust domain-specific retrieval system we've built, fueled by three distinct sources. Leveraging Medwise, a popular clinical Q&A platform, a recognized diagnostic techniques textbook, and curated medical web articles via the Brave API, we've created a rich knowledge base. Coupled with meticulous prompt engineering (thank you for the amazing tips), this significantly enhanced HealthECHO's analytical prowess.

What we learned

The journey was a great learning curve, from using Langchain to deploy our LLM app, tinkering with the new MongoDB vector search to improve knowledge retrieval, to the collaborative learning between technical and medical team members which broadened our understanding and approach to solving healthcare challenges.

What's next for HealthEcho

We envision a future where HealthEcho evolves with every interaction, incorporating feedback from different doctors into its knowledge base. This continuous learning will allow HealthECHO to ask, suggest, and learn over time, fostering a more nuanced and diverse understanding.