The problem Medvice solves

Medvice is tailored for individuals seeking reliable health information and management tools in one place, particularly those who are:
Frustrated with the inaccuracies of self-diagnosing through internet searches.
In need of a unified platform to visualize and interact with their FHIR data.
Healthcare professionals, such as doctors and nurses, who require a comprehensive, yet streamlined view of patient history for better diagnostic and treatment planning.

Features:
Medvisor: At the heart of Medvice is the Medvisor, an intelligent chatbot built on fine-tuned language model algorithms, specialized in interpreting medical queries. With the unique capability to analyze a user's FHIR data alongside current symptoms, Medvisor offers preliminary diagnoses and actionable advice. Whether it's suggesting a doctor's visit or providing tailored diet and nutrition plans, Medvisor acts as a first line of consultation, ensuring informed health decisions.

Skin Cancer Detection: Utilizing cutting-edge vision transformer technology, this tool analyzes images of skin lesions to assess the likelihood of skin cancer, offering a quick, non-invasive preliminary screening method.

Patient Information: Medvice transforms the way FHIR data is viewed by aggregating and visualizing medical records from various providers. This intuitive interface presents health information in easily understandable charts, enhancing patient understanding and engagement with their health data.

Challenges we ran into

We ran into multiple issues when trying to build our models. When we faced an 8-hour loss of our models as Hugging Face servers were completely down, we focused on researching our implementation more and transitioned from a simple foundational model for inference to a cascade model with multiple fine-tuned LLMs for increased accuracy. Limited compute and environment issues on Intel Developer Cloud led us to ensure model checkpoints were saved and data was exported continuously. Collecting data from multiple sources and putting all of it together so that it could be ingested by our text-to-query model was a challenge, which we overcame by building a parcel for the FHIR data.