Explaining Machine Learning to Fido
Machine learning is an incredibly powerful toolset that can be harnessed to automate tasks and improve decision making. I envision you nodding, sagely; I bet you’ve read many versions of this before. But do you feel like you understand machine learning well enough to explain it to your children, or for that matter, your golden retriever? For those who wouldn’t know what to say to Fido, I write this post for you. You don’t need a sophisticated math background or a computer science degree to understand it. Building a predictive model is intuitive. And I’m going to convince you of that.
Applying machine learning to bodily injury claims is one facet of what we do at Care Bridge International. Specifically, we solve supervised machine learning problems. All that means is we model problems where, in our database of longitudinal claim histories, we know the outcome. Did the claimant ultimately receive surgery? How many physical therapy visits did they have? Every problem begins like this:
But knowing whether Bill, Ted, or any other character in a 1989 American science fiction comedy film received a surgery doesn’t help us predict anything. Fortunately, we know quite a bit more. In my last post, I explained one way to try to parse the relationship between diagnoses and specific medical procedures. Having isolated diagnoses that seem to in some way predict a future surgery, we can now add them to our table.
Beyond diagnoses, we analyze our sample to identify other predictors of knee surgery. Let’s say we find that time matters: bodily injury claimants are a lot more likely to receive a treatment in proximity to their accident than otherwise. And let’s also say we find having a given procedure is a strong negative predictor of receiving it in the future. These bits and pieces of information that allow our model to derive a prediction about the outcome are called features. We add these features to our table.
Now we can utilize a learning algorithm that will enable us to predict future examples. There are many varieties of learning algorithms, but in a supervised machine learning problem, they all do the same thing: enable you to model a binary variable (like a future knee surgery, which is true or false; this is called classification) or a continuous variable (like a future PT quantity that ranges somewhere between 0 and infinity; this is called regression). A decision tree, a type of learning algorithm, works like what I’ve mocked up below.
We can train the decision tree on claimants in our database, and then later use it to make predictions on active claims. The decision tree above would predict anyone with a torn meniscus and an accident within the past 4 months is 100% likely to get surgery (because 100% of claimants in our sample got it) and anyone with an accident beyond the past 4 months or with any other injury is 0% likely to get surgery (because no claimant in our sample did). To be clear, for a decision tree, that’s it. We take a new observation for a new claimant, we follow the path of the decision tree we trained, and we arrive at a prediction.
So, if you made it this far, I hope you are thinking, that’s actually really simple! But also, I doubt the relationships that are true for these three claimants will be reliable in the future. And that second observation is correct. We will dive into it in my next post.
As always, reach out to us at Care Bridge International to talk about how we might be able to help you handle your claims more effectively through people, process, technology, and data.
To Learn more about Care Bridge International: Follow us on LinkedIn
Visit our Website
Dial Toll-Free: 888-434-9326