Medical #AI errors “can directly impact people’s health and even determine life-altering outcomes.” In @natrevbioeng, @UW #UWAllen’s @ChanwooKim, @soham_gadgil & @suinleelab emphasized the importance of transparency in medical models. #ResearchMakesAmerica

Nothing more fun than working with brilliant students @uwcse @ChanwooKim_ & @soham_gadgil on our Nature Reviews bioengineering paper!🎉We review challenges & opportunities for making medical AI trustworthy through transparency in data, models & deployment. https://t.co/ZZBdtgixVz

It was a pleasure working on this review with amazing co-authors @ChanwooKim_ and @suinleelab from @uwcse! UW News: https://t.co/8hBNDcqodT (4/4)

In this review, we examine the latest advancements in transparency for medical AI across the entire pipeline, from the training data to the model development and real-world deployment. We also provide insights to ensure transparency in emerging AI models such as LLMs. (3/n)

As AI systems become more widespread, it is essential to ensure they remain reliable, trustworthy, and explainable, especially in high-stakes domains such as healthcare. We argue that transparency should be a crucial component of medical AI development and deployment. (2/n)

Excited to share that our review paper on "Transparency of medical artificial intelligence systems" has been published in Nature Reviews Bioengineering! Paper: https://t.co/QThjxdH8uK (1/n)

🚀 We’re hiring! Multiple postdocs & a program manager to push the frontiers of explainable AI in cutting-edge biomedical research—Alzheimer’s, aging, cancer & medical AI. Start immediately. Friends, please RT🙏 Learn more: https://t.co/pAR8O6EKNP #postdocjobs #AI #Biomedicine

Congrats to @soham_gadgil on this awesome pre-print — not only does this method for vision interp use SotA techniques, but really importantly (I think) emphasizes the human-in-the-loop. So much of the value of interpretability is as a lens for looking carefully at your data

Concept is all you need! @uwcse features our @NatureMedicine paper on extracting concepts from dermatology images via foundation models, with brilliant first author @ChanwooKim_ and exceptional co-senior author @RoxanaDaneshjou at @StanfordDBDS. Details:

It was a pleasure working with co-first author Alex DeGrave along with advisors @suinleelab and @RoxanaDaneshjou! Check out our paper at:

Excited to share that our paper "Discovering mechanisms underlying AI prediction of protected attributes via data auditing" won the best paper runner up award at the Data Curation and Augmentation in Medical Imaging Workshop at CVPR 2024! (1/2)

Excited to announce the inaugural AIMBA (AI Meets Biology of Aging) 2024 meeting on May 22, 2023, 9-2pm PT, organized by @UW @NathanShockCtrs ! Featuring four keynote speakers: Profs @mariabrbic, Anne Brunet, Vadim Gladyshev, @james_y_zou. Check it out! https://t.co/O8eWzYTleZ

Excited to see our @NatureMedicine paper out today led by @ChanwooKim_ with my co-senior author @suinleelab and an amazing team! We used a dermatology foundation model to enable explainable and transparent AI - from auditing datasets to models.

Happy to share that our paper on leveraging foundation models to foster the explainability and transparency of medical AI has been published today in @NatureMedicine! Check it out here:

Happy to share that DIME has been accepted at ICLR 2024! This work was in collaboration with the amazing co-author @ianccovert and advisor @suinleelab #ICLR2024

It was a pleasure working on this project with co-first author @ianccovert and advisor @suinleelab from @uwcse! (7/7)

We generalize our approach to incorporate prior information beyond the main features. Here, we again prove that our procedure recovers a modified version of the CMI at optimality and show that this modification helps in improving performance. (6/n)

DIME is agnostic to the model used for training the network and we investigate the role of modern architectures in improving performance in DFS. Specifically, we find that for image data, our method benefits from using ViTs rather than standard CNNs (5/n)

Taking inspiration from adaptive submodular optimization, we also show how to adapt our CMI-based approach to scenarios with non-uniform feature costs while retaining high performance. (4/n)

Comparing to multiple baselines (static/dynamic feature selection methods), we find that our method (DIME) provides consistent gains across a variety of image and tabular datasets. We include real-world medical datasets where this approach is practically useful. (3/n)