Purpose

Welcome to Research to the People and Stanford Medicine's Rare Disease AI Hackathon. The purpose of this event series is bringing AI and medical experts together to build open source language models for rare diseases, and create zero-barrier access to rare disease expertise for patients, researchers and physicians. We're excited for the potential AI has to uncover new biomarkers, treatment ideas and novel links between rare diseases. As models are built and tested, this event will also explore validation methods for medical AI. Successful models will be released for further testing and validation on Hypophosphatasia.ai and EhlersDanlos.ai.

Vision

Our vision is to jumpstart an ongoing open source community for rare disease AI models.

Timeline 

See Full Calendar on AddEvent!

• April 2024 Event announced!

• Asynchronous: Teams work on models. Office hours sessions on training, fine tuning and validating models: Virtual.

• April 25th: In-person co-working in San Francisco.

• April 29th: All Hands: Social, Networking, Team Building (Online)

• TBD: Learn about Ehlers-Danlos Syndrome. Online.

• TBD: Learn about Hypophosphatasia. Online.

• See all Mentor Office Hours Sessions (Virtual).

• TBD: In-person co-working and office hours hosted by Fifty Years. 

• May 15th: Deadline to register your team to be included in the Demo Day.

• June 17th at 12.00am: Final results due.

• June 21st 2024: In-Person Demo Day in San Francisco! Open to the public.

• Post-Hackathon: Models released for Beta testing to broader patient and physician community by Research to the People. Teams will be invited to publish their work in the Research to the People Journal.

Rare Disease Focus

1. Hypophosphatasia.

2. Ehlers Danlos Syndrome.

​Datasets

Our baseline dataset are publicly available PubMed papers for HPP and EDS. Additional datasets and data sources in the works. *Teams are encouraged to use additional datasets outside of what we provide.* Special thanks to teams who are helping source new datasets to make available for this event! Data downloads available for onboarded teams. 

1. Hypophosphatasia - 814 publications.

2. Ehlers Danlos Syndrome. - 1,089 publications.

​

Problem Tracks

1. Search & Recall: Create a chat interface model for HPP and EDS scientific publications.

2. Discovery: Use additional bio datasets to create a model that can augment research for HPP and EDS, ie: new biomarkers.

​

Hackathon Events

The bulk of this hackathon will take place virtually, with teams interfacing with AI Researchers, Patients, and Medical experts through virtual and in-person (SF Bay Area) events. The hackathon will focus on fine tune training. The demo day will focus on collective interrogation of the models, especially via doctors, patients and medical experts.

Format

We're doing a rolling start, and welcoming teams to start building models ASAP. Teams can signup and access the data right away to start training their models. Virtual and in-person hackathon events and office hours will be hosted by our team and leading experts leading up to the demo day: June 21st 2024.

Over the course of these events, teams will focus on fine-tuning models with AI and medical experts. This work will culminate in a live demo day pitch session with patients, medical experts, investors and the broader AI and Rare Disease Community. ​

​

Demo Day

Our Demo Day will be open to the public. Teams will present live demos of their models and lead a discussion on their approach. This event will feature VC's and investors who are funding generative AI in healthcare and medicine.​ Note: Models developed for this Hackathon are expected to be open-source. Teams will be able to connect with VC's and Investors for other generative AI projects.

​

Recommended Tools

1. Llama 3

2. Run Llama locally with GG.ML

3. HuggingFace

4. LangChain

5. ChromaDB

6. Pinecone

7. Ray

​

Recommended Reading

1. RareBench: Can LLMs Serve as Rare Diseases Specialists?

2. The shaky foundations of large language models and foundation models for electronic health records.

3. Challenges and Applications of Large Language Models.

4. Can Generalist Foundation Models Outcompete Special-Purpose Tuning? Case Study in Medicine

5. LINC: A Neurosymbolic Approach for Logical Reasoning by Combining Language Models with First-Order Logic Provers