Machine Learning Perspective
Making ML work in high-stakes settings
Recently developed techniques in machine learning have proven adept at solving problems in computer vision, natural language processing, board games, video games, and other domains where computers had previously struggled to outperform humans.
However, deployment in high-stakes settings is still limited, for good reasons:
- Predictive models alone may be insufficient - we may care about causal questions (as in personalized medicine), not just predictive ones.
- We may require models to be robust and generalizable, able to deal with noisy, biased data or dataset shift over time.
- We may have additional requirements beyond accuracy and generalizability, such as ensuring notions of fairness or interpretability in model outputs.
We believe that healthcare is an excellent setting for studying these problems and finding rigorous solutions, which would help pave the way for safe deployment of machine learning beyond healthcare.
Bridging the gap between theory and practice
Deploying machine learning in real-world healthcare settings is challenging, and requires deep clinical expertise and intuition. However, real-world deployment is essential for not only having a positive impact on patients and clinicians, but also for surfacing non-obvious problems with existing methods.
To that end, we work with clinical collaborators to:
- Identify clinical questions where machine learning could make an impact
- Develop novel methods to overcome limitations of existing methods, and use them to solve the problem at hand
- Evaluate our solution in the context of the clinical application
We believe that publishing papers is not enough - In order to truly “field test” our methods, and make a difference in healthcare, we need to demonstrate that they are practically useful for solving real clinical problems.