When your health plan looks for an A.I.-powered solution for the SIU, here are 4 reasons deep learning should top your list.
If your health plan is looking to improve your fraud detection capabilities – by reducing false positives, finding more missed opportunities or reducing manual processes – you’re probably considering adopting more advanced technology. When exploring your options, no doubt you’ve found mentions of “artificial intelligence” in this area intriguing. We’ve written before about how you can determine if these claims are valid or not, and while we have provided some insights on what different applications of A.I. may offer to healthcare, we want to get more specific about the ideal match for fraud detection: deep learning.
So, what is deep learning, and why is it particularly suited to the complexities inherent in predicting instances of potential fraud?
The Confusion Around A.I.
With the significant challenges facing the SIU, it makes sense that progressive health plan fraud organizations would be interested in emerging technologies that harness the power of A.I. But the exaggeration of the technology in pop culture leads to many misperceptions of what it is and how it is used. We tend to think of A.I. as a sort of a ubiquitous solution to a wide variety of problems without a lot of thought to how it works.
Instead, our focus should pivot to the underlying methods that drive A.I. This wide variety of methods handles decision-making problems – primarily classification and prediction. And no application of A.I. is better at handling the complex situations presented by healthcare fraud than deep learning.
What is Deep Learning?
Deep learning is a subset of A.I.-based technology methods. It relies on advanced technologies like machine learning, neural networks and supervised/unsupervised methods to process structured and unstructured data in a way that mimics human intelligence. Deep learning methods are used for detecting financial fraud, predicting healthcare costs and outcomes, and for smart home devices translating the signal that is your voice into recognizable inputs.
We are not yet at the point where we can ask a voice assistant, “Is this claim fraud?” But the methods that allow the device to recognize the questions we ask are some of the very same methods that we can apply to fraud, waste and abuse detection in the highly complex healthcare data space. Though, at this point, it still requires a fair amount of subject matter expertise to accomplish that effectively (more on that below).
Deep learning for healthcare fraud detection: peer-to-peer provider comparisons
Fraud detection models utilize raw claims, audit data, and demographics of providers and patients. The models use many derived features that involve peer-to-peer comparison of providers:
- Overutilization or misuse of services
- Provider share of volume and claim dollars in their peer neighborhood
- Provider share of patient costs
- Discrepancies between the diagnosis and the treatment
- Patient and provider geography
- Recency of billing to the payer
The Complexity of Healthcare Fraud Detection Meets Its Match
As we think about the FWA detection analytics problem – and the benefits that this broad class of deep learning methods and technology bring to that problem – it may be helpful to contrast healthcare FWA against other industries that have had success with the application of deep learning analytics to the detection problem. As you might recall from our previous article on the topic, the comparison that most often comes to mind is credit card fraud, despite a huge difference between the two: complexity.
Challenge: Very little structured, historical claims data labeled “fraud”
Even if financial institutions didn’t detect credit card fraud at the moment of the transaction, the fraudulent activity would still be found upon consumer evaluation of the monthly statement. That confirmation data, along with the exponentially larger numbers of credit card transactions vs. healthcare claims, provides financial institutions with reams of structured, labeled historical data to train fraud models.
On the other hand, even upon close inspection of EOBs provided to consumers, these documents lack the specificity and understandability to root out healthcare fraud at this level. This difference in the historical data leads to the fundamental problem with advanced analytics, which is the problem of making correct classifications. How accurately can you identify fraud as a fraud, and a non-fraud as a non-fraud?
Solution: Deep learning can handle vast amounts of complex, unlabeled, unstructured data
While some information useful for fraud detection comes in the form of relatively structured claims, most data must be mined from free text clinical information. It’s incredibly inefficient for investigators to read through pages of information and reference several applications while looking for the needle in the haystack that is potential fraud. One of the best use cases for deep learning is taking unstructured clinical documents (notes, transcripts) and converting it to structured concepts with actionable information.
Moreover, the unsupervised deep learning acts in a way human expertise cannot. We know there’s probably fraud in the data, yet we’re not sure what it might look like when we find it. Which leaves us in a position of being unable to say, “I’ll know it when I see it.” But rather, “I’ll say and do it after I know that I’ve seen it.” Making these connections is what deep learning excels at, and it performs even better at this task with massive amounts of relevant data – beyond what limited historical claims information can provide.
Challenge: False positives are expensive to handle
Even with enviable amounts of structured, labeled data, credit card fraud detection still averages a 90% false positive rate that no health plan could reasonably accommodate. But, with financial institutions’ largely hands-off approach to resolution, they are easily able to keep their costs low.
SIU teams are far too under-resourced to sift through that many red herrings in search of legitimate fraud. Plus, in the effort to prove or disprove a fraud lead, the investigator likely will need to request additional information from the provider at some point – and risk unwanted abrasion in the process. Health plans understandably want to maintain these valuable relationships, and the vast majority of providers submit legitimate claims with the best of intentions, so false positives are a business risk for healthcare more so than other industries.
Solution: Deep learning processes data in multiple layers and dimensions to fine tune results
To offset the risk of false positives, healthcare fraud detection solutions must go beyond the “outlier analytics” version of deep learning that financial institutions can leverage. It’s a tradeoff between sensitivity (catching more potential fraud) and specificity (catching only very specific fraud).
We don’t have a way to directly measure the accuracy of the analytics that we apply in unsupervised learning methods, so we have to use some indirect measures to make sure that the algorithms that we apply are accurate enough to not exacerbate the cost considerations of false positives. Deep learning is able to transform the space of the underlying data and go beyond provider anomalies through multiple layers of processing. As a result, you can better balance between false positives and false negatives.
Challenge: Impractical to develop rules for all possible scenarios
If your SIU has been working to mitigate fraud for any length of time, you likely have a trusty rules-based system at your disposal. These systems usually come pre-populated with a number of fraud-focused rules or concepts that are run against incoming batches of claims. You may also have a data science organization – or access to resources shared with IT or other areas of payment integrity – that work to augment this native functionality with your own proprietary rules.
This competency is foundational technology essential to any healthcare fraud organization. But the process has its limitations – especially if your data science resources are limited. Theoretically, we can build as many rules as there are examples of fraud, waste, and abuse – but at a certain point, that becomes impractical. Additionally, FWA schemes tend to stay a step ahead of the rules we build, despite our best efforts. The unfortunate result of an analytics-only system is one that can only surface known schemes.
Solution: Deep learning evolves to “get smarter” over time without explicit programming
To the extent that we can build a base of rules is always useful, but advanced analytics based on applications of A.I. are what modern fraud data science organizations need to accelerate their capabilities. Fraud detection methods based on deep learning can, for instance, scrutinize links between multiple providers to see if an unethical or illegal activity associated with one provider may also be practiced by another. These advanced analytics can also explore demographic, geographic and other data about members and providers to double check or justify claims that may raise a red flag – all without being explicitly programmed to do so.
Deep learning fraud detection solutions are also not limited to specific points in time. Historical information augments and informs the new data arriving and updating daily to get smarter over time. As a result, you do not need to update models every time new information comes in to identify short-term spikes and longer, more subtle trends in the data. This is where advanced technology excels for the SIU.
Challenge: Healthcare fraud too complex to eliminate the human element
Like all areas of health plan cost containment, fraud organizations are being charged with doing more with less. But with the high stakes of a fraud investigation and the complexity of fraud data, decisions often come down to value judgments. How can you make the most of your expert resources so they know which leads to focus on and why?
Solution: Deep learning can leverage domain expertise
Deep learning won’t replace your fraud team. Rather, it informs them by reducing the need for manual, repetitive work. It gives your team the ability to prioritize leads, focus on strategic analysis and improve accuracy. Moreover, the domain expertise that investigators and fraud organizations have spent years developing is an excellent input to more finely calibrate the results of algorithms based on deep learning methods. When investigators work alongside a deep learning system, their reviewing and labeling results improves and optimizes the analytics.
Deep Learning is Key to Pareo Fraud: Detection
For all of these reasons, deep learning is the core application of A.I. selected by ClarisHealth to power our Pareo Fraud: Detection solution. It’s able to dig deeper and wider into your data to find those very subtle changes and nuances in provider and member behavior that may indicate fraud, waste or abuse that even expert investigators may not be able to find on their own. Our 4-tier scoring methodology – configurable by health plan priority – allows the models to self-learn and evolve so that the false positive rate reduces with each round of provider scoring and data ingestion.
Pareo Fraud: Detection is also able to transcend the explainability issue that creates confusion in many A.I.-based systems. It was designed to address this “black box” gap through a specialized Super Score output – a separate deep learning method – that provides insights into results to engender trust by users. And because it’s built onto the Pareo platform, it creates an end-to-end payment integrity solution that fully eliminates gaps in a health plan’s cost containment continuum, from prepay to post-pay.
Now’s the time for total payment integrity
See the ClarisHealth 360-degree solution for total payment integrity in action.