When it comes to artificial intelligence (AI), you can ask all the right questions, hire all the right data scientists, train all the right machine-learning models, and still end up with challenges. No matter what you do, you're going to run into unknowns when you release AI models into the wild.
In this article, two data-science experts from Optum® share the lessons they've learned for how to overcome the challenges that come with implementing AI in your business units.
Four steps to a successful AI project
Before you launch any data-science project, we recommend using this simple four-step outline. This is adopted from years of working with technology and data projects.
- What are you working on and why are you working on it? (Can you quantify its value?)
- What are you going to predict? (Is the result enough to base decisions on?)
- How will you use that prediction? (What is the practical business use?)
- How are you going to change health care using this? (These technologies are impactful, and we want to make sure that we're working on impactful projects.)
If you can answer these four questions, you're going to have a fairly successful project. With every project that goes off the rails, we ask, “Where's the four-step outline?”
It's also important to get IT involved early, because they are the experts at change management, rolling out releases, training users, and integrating systems. Data science is different from IT, even though it can look similar. But when the data scientists are done with the modeling, those models still have to be integrated with existing systems. And the four steps help everyone know what they need from each other.
Implementing AI will create problems
Data is constantly changing, moving, drifting and morphing. If you train an AI model on test data that is two or three years old, you can't put it into production and expect the model to do everything right.
Sometimes the solution to this is very simple and basic: We can launch our model, track its performance over time, and set some upper and lower bounds. When it goes outside those bounds, we know we'll need to retrain it. But that's the best-case scenario; often the solution is more complicated.
Health care has frequent changes in regulations, policies, and care guidelines. New entities enter the marketplace often. You might have to go back to the drawing board and rework the fundamental assumptions and functions of the model. Also, patterns and preferences in health care and lifestyle change over time. So a model is not a destination, it's a journey. You'll need to keep revisiting, refining, tuning and adjusting it as you go.
Flexibility, agility, tenacity
These implementation issues are not easy problems. They require a tremendous amount of political and organizational skill. You will run into technical challenges, process challenges, and data challenges. You will encounter any number of road blocks along the way, and you will need to be comfortable with the fact that there may be a lot of bumps to get over.
But there's almost always a way to get from where you are to where you want to go. We've started a lot of projects where the business says, “This should not work, that will never work,” and the team comes back and says, “Yeah, that worked.”
The way you create value is by being agile, tenacious, and flexible. You need to keep getting up every morning, driving forward, and being willing to adjust to what you learn, because every day you're going to learn something new.
Answers you might not be looking for
Here's an example of a problem that we needed to look at differently. AI technologies need a lot of data to find patterns, but one area that has very little data is rare diseases. We had a theory that people with a rare disease would probably exhibit the same patterns: going from specialist to specialist, searching for answers.
Our thought was, “How do we identify them and offer them a differentiated service?” We realized we can send them to a call center that specializes in rare-disease cases, so they talk to someone who is trained in empathy and has extra resources, to give them a better experience during a very difficult time. We might not be able to solve the rare-disease problem, but at least there will be somebody on the line who understands what they're going through and can help them, so they can at least walk away thinking, “With all of this horrible stuff going on, Optum listened to me.”
So in a case where we have very little data and there's no easy answer, we found an answer — just not the one we were looking for — by thinking differently about the data that we do have.
Change hearts to change minds
When you start to implement AI in your business units, the biggest challenges will be the change in business processes and helping folks get comfortable with a different way of making decisions. How do we change the organization to be more data-driven, data-centric, and data-savvy in terms of what we do? How do we get the organization to a place where the investments we make are based on data, analysis and rational debate?
The more we look at it, the more we see the importance of community and culture. In a sense, what we're doing is a culture initiative. And the way to change the culture is to build up a community. Very few people change their behavior on the basis of an algorithm, rather because of a shared ideal.
We can't tell you the number of people who've said the most valuable parts of our Data Science University training weren't necessarily the equations and the formulas. The real value was the shared belief in the power of curiosity, creativity, tenacity, humility, integrity and compassion, and a commitment to serve others, whether that's business, IT, providers or patients.
Learn more about artificial intelligence, data science, and how Optum is improving health care with these technologies at optum.com/cio.
This is the third article in the Optum "AI talent/skillset" series. Read the first article for practical AI advice for top leaders. Read the second article for insight into hiring vs. training.
© Optum, Inc. All rights reserved.