Lessons From Evaluating and Debugging Healthcare AI in Deployment

Explore the challenges and opportunities in translating trustworthy AI from research into healthcare deployment in this Stanford seminar. Delve into insights gained from conducting real-time AI trials and analyzing FDA-approved medical AI systems. Learn about data curation techniques, quantifying data contributions to model success or biases, and continuous real-time testing and explanation of model mistakes. Discover strategies for designing AI to optimize clinician performance and human-AI interactions. Examine topics such as improving clinical trial efficiency, addressing performance issues in dermatology AI, identifying ethnic stereotypes in language models, and using Data Shapley Value for fairness improvement and dataset bias detection. Investigate neuron-level analysis for understanding model behavior, conceptual explanation of mistakes, and natural language model editing to reduce bias. Gain valuable takeaways on the shifting challenges from model training to evaluation and monitoring in healthcare AI deployment.

Introduction
Al make clinical trials more efficient
Why did the Derm Al performance crater?
Language model captures ethnic stereotypes
Two Muslims walked into...
Data used to train dermatology Al
Data Shapley Value
Dermatology classification
Shapley value identifies mis-annotations
Data Shapley improves fairness
Auditing ML data w/ data Shapley
Understanding what the network is doing
Sparse neurons responsible for prediction
Neuron Shapley identifies dataset bias
Model repair by removing bias neurons
Why did the model make this mistake?
Conceptual explanation of mistakes Mistakes made by the model
Natural language model editing reduces bias
Takeaways: challenge shifts from model training to evaluation and monitoring