Ai integrations

[adamkells]

Addresses #27

[jenniferjiangkells]

lgtm 👍

@adamkells I think you have the right idea to have a unified API that allows users to load commonly used medical AI models with the whole pipeline - loading, pre-processing, running the model, post-processing that is easily extensible.

Personally I find that I sink a lot of time into just making models run since some of the medical models we have in academia are just SO poorly documented so an API like this will make it easier to build re-usable pipelines for different use cases.

In the future we can hopefully make evaluations and benchmarking of these integrations super simple too. ✨

[jenniferjiangkells]

Lots of general models and foundation models will do the job but here are some domain specific ones. Might want to divide them up by the kinds of tasks they are suitable for, prioritising the use cases we have currently implemented.

llms

• meditron https://huggingface.co/epfl-llm/meditron-70b
• med-gemini
• med-PaLM

transformers

• ClinicalBERT https://github.com/uf-hobi-informatics-lab/ClinicalTransformerNER (not sure if on HF)
• GatorTron https://huggingface.co/UFNLP/gatortron-base/tree/main
• MedPhi-2 https://huggingface.co/bluesky333/medphi2
• EHRShot https://huggingface.co/StanfordShahLab/clmbr-t-base

spacy

• MedSpacy https://github.com/medspacy/medspacy
• SciSpacy https://allenai.github.io/scispacy/
• med7 https://huggingface.co/kormilitzin/en_core_med7_lg (yeah I know)
• medcat https://github.com/CogStack/MedCAT

[jenniferjiangkells]

https://github.com/AI-in-Health/MedLLMsPracticalGuide
https://github.com/burglarhobbit/Awesome-Medical-Large-Language-Models