I was on the faculty of the AIMed Conference, which was just held in Laguna Beach California. This gave me the opportunity to discuss the value of different approaches to create predictive models for analyses of patient monitoring data. Specifically, my overview focused on the differences among:
- White box models that incorporate well understood relationships of human physiology
- Black box machine learning models that are purely extracted from data
- The approaches in between, hence the title of my talk “Fifty Shaded of Models”
I used a famous quote from a Secretary of Defense from the recent past to illustrate how this models capture the world to predict consequences. The premise is that models can incorporate:
- “Known knowns”—completely deterministic relationships between action and consequence
- “Known unknowns”—relationships that are uncertain and need probabilistic models
- “Unknown unknowns”—unknown effects that lead to unknown consequence, which can only be learned by Big Data methods.
At Etiometry we believe that there is great value in creating predictive models from the known physiologic relationship. This has helped us produce algorithms that incorporate existing knowledge, are capable of distinguishing causal relationship, and can scale to present increasingly accurate description of human biological processes and incorporate more personalized information such as genetics.
However, we also believe that in our quest to continuously improve our predictive models there is a great value to reduce the “unknown unknowns” and discover new relationships and predict new consequences.
In this respect, we think about our scalable physiology-based modeling framework to algorithms as the right basis to provide context to machine learning that can push the boundary of known physiologic relationships and take full advantage of the massive amount of patient data we are collecting. If you’d like the slide or more information on my presentation, please send an email to email@example.com.