Over the last two weeks, I heard two interesting perspectives on machine learning that illustrate the complexity associated with tools touting this functionality:
- From a CMO: “We purchased a machine learning tool that was supposed to predict the likelihood of sepsis, but it was complete garbage. How do we know your system works?”
- From a VP of Periop: “We are too busy to put the puzzle pieces together. We want a machine learning system that makes recommendations for us.”
These two statements represent a Catch-22 for a company like Hospital IQ. On one hand, there is a desire to have a “Simple button,” but at the same time users often don’t trust a simple button because they don’t know how it works. When you buy a car, you don’t go to the dealership and say, “I want the blue one.” You take the car for a spin, you review the specs, you haggle on price and financing. This model should be no different for machine learning solutions. There are two ways you can metaphorically “kick the tires” for these types of solutions:
- Review the specs: Bring in the stakeholders who are knowledgeable about the subject matter and have them drill into the details of the recommendations to give them a sniff test
- Give it a test drive: Test the software in a limited setting and confirm the information and recommendations provided are accurate
In my experience, these approaches are used in sequence. No one is going to test drive a machine learning solution until they are sufficiently convinced it will work. The trickiest part is to provide the right level of detail to make machine learning solutions BOTH simple and understandable (and therefore, trustworthy). We want to avoid the “black box.” If I try to explain multiple linear regression models to a hospital executive, 9 times out of 10 their eyes will glaze over. But if I explain that historically a specific surgeon books 80% of her OR block time 23 days ahead of time, she has only booked 40% of her time for an upcoming surgery date, and therefore make the recommendation that she release 40% of her time, that should resonate more deeply. It’s in understandable and relatable terms.
Your hospital should approach every machine learning solution with Ronald Reagan’s “trust with verification” mentality. No machine learning solution can be expected to be the silver bullet for your issues and there are definitely snake oil salesmen out there inflating what machine learning can and should do. You should not look for the simplest machine learning solution, but rather the one that is simple enough to trust.