Everything you need to know about Robotic Process Automation, Machine Learning, & Artificial Intelligence when selecting your next healthcare solution.
December 28th, 2020
Over the past several years, a few terms have taken the healthcare industry by storm. First, it was robotic process automation or RPA. As soon as everyone came to terms with the idea that they needed to look into this "emerging" technology, it was overshadowed by artificial intelligence (AI). Suddenly, RPA was inferior technology, and anyone that's anyone needed AI. To add to all the confusion, we start peppering in the term "machine learning" (ML).
The truth is, these are all mostly buzzword marketing terms.
Robotic process automation has been around for eons, but you would have called it "scripting." Most of what is labeled as artificial intelligence is not artificially intelligent at all, and many products labeled as RPA aren't. They're prepackaged solutions built on RPA to complete a very specific workflow. So far, I have yet to find someone that can explain machine learning to me in a succinct way.
So, what is a person making or recommending purchasing decisions to do? You know you need to make investments that will boost productivity and efficiency. You know you don't want to invest in a technology today that will be obsolete tomorrow, but you also can't throw money at an emerging and largely experimental technology years away from maturation.
To help answer some of these questions, I've gone outside the healthcare industry for definitions. I've stepped outside the marketing brochures and delved into the scientific research. What I've found has been enlightening and just might help cut through the noise.
RPA, ML, and AI Defined and Described
Robotic Process Automation
CIO.com defines robotic process automation as, "an application of technology, governed by business logic and structured inputs, aimed at automating business processes (Boulton, 2018)."
What strikes me most about robotic process automation isn't how we define it. It's that if you listen to the hype, it's a brand-new technology with only a handful of suppliers. This couldn't be further from the truth. Databound's automation tool, EMUE, has been around since 1999. Its predecessor, OLIE, was around long before that!
No, robotic process automation is not a new technology or an emerging technology in healthcare. People have been automating healthcare business processes since the mainframe days.
You hear "machine learning" less often than robotic process automation or artificial intelligence. I attribute that to the fact that it doesn't sound cool and hip. Here's what I find funny. Vendors and marketing firms are incorrectly applying the moniker "artificial intelligence" to robotic process automation. But when they describe the artificial intelligence their solutions don't have; they're talking about machine learning! No wonder buyers are confused and skeptical.
Machine learning is, "the study of computer algorithms that allow computer programs to automatically improve through experience (Mitchell, 1997)."
Whoa. That's a lot of words. Let's see if we can boil that down a bit.
Tom Mitchell is saying that machine learning means that the computer program can learn from its own experience. When I think of this, I think of those little toys that bounce of walls. They run across the floor, hit a wall, back up, turn, and start running again. The difference, of course, is that the toy will continue to hit the same wall. For true machine learning, the toy would have learned the wall's location and avoided it evermore.
When folks say something like, "suppose you send out thousands of claims with a diagnosis code, and all of the ones denied were because the claim lacked an authorization code. AI would see that pattern and learn to make sure there is an authorization code before submitting the claim," they're not talking about AI. They're talking about machine learning. It can be a very powerful tool.
Now here's where it gets tricky. Artificial intelligence is "the science and engineering of making computers behave in ways that, until recently, we thought required human intelligence (High, 2017)." Let’s unpack what that means in comparison to RPA. Ages ago, if you wanted to add coded comments to a claim, a human had to do it. A computer wouldn’t have known how to determine if the comment was needed, which claim to add it to, or which coded comment to add. RPA can now do that, which means RPA can behave in ways we thought required a human. So, does that mean RPA is AI? Well- technically….yes. But, would that fit anyone’s definition of AI today? No. It likely would not.
Artificial intelligence isn’t a technology. It’s a concept achieved with a variety of technologies. What is considered “until recently” today will expire tomorrow, thereby changing the criteria needed to call something AI. Many things have fallen into the AI category like- RPA. Machine learning. Predictive analytics. Deep learning, neural networks. AI at it’s core, is a moving target.
Decades ago, playing battleship required two humans. Now, you can play against a computer.
The definition of AI is structured so that any advance in AI eliminates that advance’s ability to be considered AI. This was intentional because it means that AI will never be done and insists that it continues to evolve.
I have to admit, though. Seeing “AI” on a marketing brochure sure looks slicker than “machine learning.” Look, Ma, my very own C-3PO!
At this point, you might be thinking, “Ok, I understand the definitions but, so what? How does that help?” That’s a very good point. It only helps as we start to describe which technology you should be considering for which application. So, let’s do that. Let’s also talk about why you would use a certain technology instead of the other.
Automated processes have been given the name “bots”. Each little thing that goes out and does work for you is a bot. Think of a bot as a low-ranking soldier. They are given a task to complete and a set of instructions to complete it. They do not have the authority to deviate from the plan. If decisions need to be made, there are defined decisions to choose from and criteria for making the selection. They are good at getting a ton of grunt work done so that others can focus on more complex decision making. That’s RPA.
RPA is your grunt force. It does all the stuff your staff hates doing because it’s boring, time consuming, repetitive, and requires no thought. You hate seeing your staff doing it because you have a mental list of all the other stuff that could be getting done if they weren’t sitting there hitting the same keys on the keyboard over and over again.
It makes me think of a quote a friend of mine at Moffitt Cancer Center uses when she sees her employees working through highly repetitive tasks. “Hey! No clicking! Use EMUE!” She even has a sign on her office window with a mouse and a line through it that says, “NO CLICKING!”
Consider things like adding coded comments to thousands of encounters. Applying updated insurance information to thousands of patient records. Clearing out thousands of $0 dollar or low balance accounts from A/R. Re-registering recurring treatment series. Cleaning up voided benefit orders. Writing accounts off to bad-debt. All of these things:
- Have a consistent data source
- Have structure data
- Require minimal decision making
- Have an obvious set of options and criteria for decisions being made (If it's $0, clear it out. If it’s $0-$2,000, write it off to bad debt. If it’s $2,000-$5,000, add it to a list for review by A/R staff. If it’s over $5,000 and less than 120 days, submit it to collections. Etc)
- Are manual in nature
- Are data intense
- Are repetitive
This is where RPA shines. Some may say, “That sounds pretty limited. It seems like I could get more done with ML.” I’d say you wouldn’t necessarily get more of the same done with ML. You’d get different things done with ML. Much like it would be a horrible application of resources to drive a tractor-trailer on a snack run, it would be a horrible application of resources to use ML on processes that are straight-forward and well-defined.
Now you might say, “But if I have ML, why not?” Well, let’s go back to our tractor-trailer analogy. Sure, you could use a tractor-trailer for snack runs, but the fuel would cost 4x. Maintaining the vehicle is both more expensive and more intense. You need specialized equipment to work on a tractor-trailer, and you need more storage to park it. The driver requires a different skillset and licensing than one driving a car. Overall, it sucks up available money and talent to drive the tractor-trailer that could be used on a more appropriate use. The same logic applies to RPA vs. ML. It costs more, requires different hardware, and requires a higher skill set to operate. It’s merely a bad use of resources.
Machine Learning Applications
Machine learning might also use something that could be called “bots”, but these bots will be much more complex. Remember when the streaming music service, Pandora, came out? I do. I was pumped! Here’s a thing that can provide a variety of music I actually like. Pandora had a destination in mind. They wanted to give their platform the ability to recognize a user’s musical preferences and play songs that the user will enjoy. It’s more than, “If Mike likes a band in the 70’s rock category, play more songs from the 70’s rock category.” That would have been a good case for RPA. Pandora looks at different data like rhythm, genre, tone, feeling, and more. It then sifts through thousands, if not millions, of songs for similarities to decide what to play next. That made machine learning necessary.
The application of machine learning is more appropriate for analysis than for labor. For example, I will rely on someone smarter than me on the subject, as my experience is largely in RPA:
“In a simple example, if you load a machine learning program with a considerably large dataset of x-ray pictures along with their description (symptoms, items to consider, and others) it ought to have the capacity to assist (or perhaps automatize) the data analysis of x-ray pictures later on (Iriondo, 2018).”
We learn that machine learning is appropriate when you don’t know the right course of action to take and need a considerably large number of scenarios and outcomes evaluated to identify the contributing factors of the desired outcome. That evaluation will lead to a very well-informed decision that, by all measures, would appear to use intuition.
ML requires special care in the healthcare environment. If Pandora plays a song I don’t like, I get mildly irritated and skip it. If ML makes a faulty diagnosis, well…..you get the point.
I promised to offer some help in determining when to apply a specific technology and determine whether the hype meets your needs, so here goes.
When looking at what you want to accomplish, ask yourself this, “Do I want to get a bunch of work done and free up my staff for other stuff, or do I want help determining where my processes are broken and how to fix them?” Chances are, you will answer yes to both. That means you need one of each.
When looking at RPA vendors, you need to make sure you’re evaluating end-to-end tools, not one-process solutions. Many vendors are labeling specific solutions, like claims follow-up, as RPA. It’s not RPA. You can’t apply it to whatever you want, and you’re limited to claims follow-up. So, make sure you’re buying an automation tool you can adapt to your needs.
When trying to make sure the hype is matching reality in AI or ML investments, there’s one question you need to ask your vendor. “Can this application be trusted to evaluate large quantities of information, see patterns and trends, and make its own decisions while I’m 100 miles away from my computer keyboard?”
If the answer is yes, make sure you follow that up with, “Ok. Show me proof beyond brochures. I want to see the application do it real-time in the environment I’m working in.”