With the onset of ICD-10 came exceptional concern over what would happen to denials – and with good reason.
Denials are already one of the most persistent problems in the revenue cycle and anything that could impact the delicate machinery of denials management is worthy of trepidation.
Consider this for a moment: denials performance can swing a hospital from profitability into loss. For the average hospital in Becker's Top 100, denials represent $159 million in revenue at a 3.8% rate of denials.
Even accounting for re-work, the amount of written off denials exceeds $100 million for the average Top 100 hospital.
This is precisely the reason why everyone from CEOs to RCM managers are dealing with ICD-10. It would appear that the preliminary work has been worthwhile – as early data indicates less of a catastrophe and more of a bump in the road for the RCM teams.
What is going to be lost in the Y2K comparisons, however, is that ICD will have an effect, it will just be harder to discern and more difficult to quantify.
Let me explain.
ICD-10's combinatorial complexity hardens the problem.
By that I mean that it makes denials as a whole more intractable – almost like a layer of concrete. To break through it will require a machine, but unlike a jackhammer – this one will be software.
Combinatorial Complexity and ICD-10
Dealing with denials is already hard, but now teams have to deal with a stunning increase in the number of possible code combinations. And when I say stunning, I mean the number of atoms in the universe stunning.
To put this in context, if your team is well acquainted with 35% of the 14,000+ codes under ICD-9, that would represent around 4,900 codes.
The number of possible combinations within those codes is too big to fit on the page. It has 1,475 zeros.
Under ICD-10, with the same expectation for 35%, the team would need to have a grasp of 23,800 codes. The number of possible combinations for that number far exceeds the estimated number of atoms in the universe. The number representing the zeros barely fits on the page.
So why hasn't the ICD world ended?
Well for one, many of those new codes are pretty specific. "Hurt at the Opera" codes, "bitten by cow" codes, etc.
Secondly, if trained properly, good coders can keep pace with the combinatorial explosion. It is new information, but if they can master it or use new search technologies, the prognosis is pretty good for keeping things as they were.
The problem is "keeping things as they were" isn't the charge of the denials team. Their job is to reduce denials and their job, courtesy of a layer of concrete, just got a lot harder.
Even today, with the tools at their disposal, most denials teams average 3.8% denials. The average team achieves recovery on about 35% of those, leaving 65% untouched.
Those go untouched, because to the denials teams, they look like individual claims. At a cost of $1,000 per claim, there are no economics that work for it to make sense to go after these claims.
What they go after is what their tools can "see" and that is generally 2-4 factor problems – a specific diagnosis at a certain facility with a private payer and this HCPS code.
These teams attack what they can see and fix – the biggest patterns. That is the 35%, and its hard work.
The problem is the 65% where patterns are more complex. Where 7-8 factors prevail. Take this example:
We have a problem with pre cert denials on outpatient MRIs with contrast when a diagnosis of sciatica is billed with one of these two HCPCS codes, the facility is not fully integrated into our network of providers, the payer is private, and the employee responsible for obtaining the pre cert has been with us for less than 6 months.
That is a real example.
More importantly, it is not one claim. There are hundreds of claims in that group, worth hundreds of thousands of dollars.
To the tools in the market today, however, it looks like a singleton, and therefore not worth the effort.
That is the problem under ICD-9.
The odds of finding the patterns under ICD-10 are exponentially harder.
So while the problem of net denials doesn't get harder, the problem of chipping away at denials does get harder.
The answer, the only answer, in this environment is machine intelligence. Machine intelligence finds the mid-level pattern, like the one described above, worth hundreds of thousands of dollars, but effectively invisible.
A mid-level pattern is both specific, actionable and large enough to merit interest. It has the attributes and specificity outlined above.
That is the kind of answer you can act on, and can act on with a nuanced and specific intervention.
It is large enough to matter, it is actionable enough to matter, it is an answer that embraces the complexity of claims data.
Why can't you get to these rich descriptions of patterns in the middle with traditional analytics?
Well for starters, traditional methods start with people asking questions.
To get to a mid-level pattern that you can action, you have to ask a very specific question. From our example above you would need to address denial type (pre cert), procedure (MRIs), account class (outpatient), procedure codes on claim (HCPCS), facility type (non-integrated), payer type (private), and longevity of particular staff members.
That's seven variables.
If you code, imagine that complexity of that query. If you don't code imagine that string of pivot tables.
But why those seven variables? Why not some other seven?
Maybe it's patient status, referring provider's department, existence of an ECI code on the claim, and patient's age.
That's a different query.
How many of these queries can you write? How long can you spend pivoting?
This is combinatorial complexity.
If you're capturing one hundred features of every claim (and actually that's a really low estimate), how many combinations of seven features are there?
Sixteen billion. That's sixteen billion questions you can ask by doing 7-fold combinations of those one hundred features.
What if you needed a ten-feature combination to characterize the pattern?
Enter the Machines
To crack the ICD-10 veneer, you have to build a learning system that is designed to find the answers by detecting all of the relationships associated with the data – one that finds answers without asking questions.
To be clear, we are not suggesting that software eliminates people from the workflow and drives denials to zero, that is unrealistic.
What we are saying, however, is with the advent of technologies like machine intelligence, one can leverage software and computing power to source the answers in a fraction of the time it would take traditional technologies and fraction of a fraction of what it would take a person.
The speed and comprehensiveness of machine intelligence put providers on the path to zero, but it doesn't happen automatically on day one.
A machine intelligence framework reveals the subtle, interconnected elements of a particular payer's application, misapplication or non-application of their own rules. Machines have the capacity to relentlessly uncover patterns in payer behavior, independent of the actual "rules" - focusing on outcomes, behaviors and subtlety – dispensing with concepts such as "should" and "intent."
From Discovery to Resolution
Discovery is just the first step, however. What is important is what happens next.
First, you identify drivers of these mid-level patterns group by group.
Second, you prioritize work queues by identifying which group a denial class belongs to. This can be streamlined by giving the person working the work queue a disposition – an expectation of reason for the denial.
Third, you inform upstream process changes by making them precisely targeted and nuanced. You change what needs to be changed, and minimize disruption in the revenue cycle with precise solutions.
Now, the domain experts evaluate groups of similarly denied claims and suggests process improvements using the underlying characteristics of each of these groups.
By identifying, modifying, and fixing processes upstream, hospitals can minimize the number of claims that are rejected or denied.
The Net Net
Denials are persistent, chronic issues in the revenue cycle and solving them just went from hard, to really, really hard with the addition of the new ICD-10 codes.
To meet this challenge, providers will need to understand the multi-factor problems that define the majority of denials and be able to look at them as a class, while contending with the problem that these denials look like individual claims to most solutions being used today.
The multi-factor, mid-level patterns that machine intelligence can identify in the complexity presented by ICD-10 represent actionable groups of claims. These patterns hold the key to both revenue recovery and upstream process changes that attack denials at the source, and will allow hospitals to truly reduce denials, instead of "keeping things as they were" with ICD-9.
Prashanth is Head of Product, Healthcare where he is charged with leading the definition and build out of Ayasdi's healthcare product portfolio for the provider and payer segments. Prior to joining the company, Prashanth was the senior director of health-care product strategy at Oracle. He brings over 20 years of strategic technology experience to the role, having served as a senior scientist with Proctor & Gamble. Prashanth earned his Ph.D in electrical and computer engineering from the University of Cincinnati with a focus on medical imaging.
The views, opinions and positions expressed within these guest posts are those of the author alone and do not represent those of Becker's Hospital Review/Becker's Healthcare. The accuracy, completeness and validity of any statements made within this article are not guaranteed. We accept no liability for any errors, omissions or representations. The copyright of this content belongs to the author and any liability with regards to infringement of intellectual property rights remains with them.