“If only I knew then what I know now.” Who among us hasn’t said that at some time or another, either as we reflect back on things we did or on actions we should have taken but didn’t?
While that thought tends to be focused more on our personal lives, and often forms the motivation for attending high school reunions, it can definitely be applied to the standard approach to automating revenue recovery as well.
Currently most organizations rely on a rules-based system which looks for binary relationships in the form of “if A and B then C.” A simple example is “if a patient is having knee replacement surgery and there is a charge for putting an artificial joint in place there should also be a charge for the artificial joint itself.” Failure to find a charge for the artificial joint immediately triggers an alert so a correction can be made and the provider can be reimbursed in full.
This is where the “if I knew then” scenario comes in. Whoever is writing the rules must develop hypotheses around all the revenue recovery possibilities so rules can be written against them. If someone doesn’t postulate an issue a rule can’t be written against it and revenue is lost.
The problem is organizations often don’t know what they don’t know, so if an unknown issue is also a common one, revenue loss can go on for months or even years before it is discovered.
The other challenge the use of a rules-based system creates is a delay in getting the revenue recovery system in place. Since there is a need and a desire for all of the work to be performed up-front to ensure its thoroughness, development time is often measured in months or years. Every second the technology is delayed as more issues are being hypothesized and more rules written, the level of potential revenue loss increases.
Predictive modeling circumvents the rules
To overcome these limitations, healthcare organizations have begun experimenting with predictive modeling, which uses machine learning/artificial intelligence and data mining to extract patterns and trends from the organization’s historical data in order to identify missing revenue opportunities. These models typically incorporate many disparate elements such as the type of visit, admit and discharge dates, diagnosis and procedure coding, the admit sources, charges, reimbursements, and costs.
The best predictive modeling technologies also pull in physician data to take the way physicians practice medicine into account. Going back to the knee surgery example, specific orthopedic surgeons often use a particular set of devices. By knowing which physician is providing care to a patient on a given encounter, predictive modeling can accurately identify the common drugs, devices and procedures that the physician typically uses. So, if any of these commonly used items are missing when the physician is treating a specific patient, these are captured and the associated revenue is recovered.
Once all of the historic data has been collected, data scientists can begin building predictive models that identify the likelihood of certain events taking place. Examples include a charge being missing around a particular claim, or whether a claim will be denied based on similar past claims, or whether a patient will pay his or her portion of a bill, or whether a patient will re-admit within 30 days, which triggers a penalty for the hospital.
These models are built using statistical algorithms that look to find the correlations between the thousands of attributes that describe a patient encounter and the event of interest, such as whether the patient will pay their portion of the bill. Different algorithms look for different types of correlations.
The major difference is that these algorithms are capable of learning as they gather more “experience” with the data. By identifying soft trends and patterns rather than following hard rules they can make determinations such as charges for insertion of an artificial knee are normally associated with a charge for the device commonly used by the physician on patients who are younger than 55. If a device charge is missing, the model will create an alert. No one had to tell the model to look for this issue (based on the historic pattern of treatment that this physician provides to younger artificial knee patients); it was able to deduce the probability of an issue based on the norms within the organization.
Gaining consensus
There are literally hundreds of algorithms that can be used to build predictive models, each of which look for different types of correlations. Incorporating multiple algorithms into each predictive model helps increase true positive rates and reduce false positives by capturing different correlation types within the source data and then reviewing which predictions agree. This technique is known as consensus predictions.
Say three predictive models with different algorithms are being applied to a particular problem. If all three agree there is an issue, the organization should have high confidence in the predictions. Even a consensus of two out of three is often sufficient to justify further investigation.
Using different models not only helps address the current issue; the results can be fed back into the system to improve accuracy for the future.
A place for rules
Predictive modeling’s strength – its ability to identify granular patterns based on historic data – also exposes a weakness: it identifies what is outside the norm. So, if the norm is incorrect, such as a hospital has never charged for an artificial knee in knee surgeries, the model will recognize this gaping revenue hole as normal and won’t identify it as an issue.
That is why it’s important to have a rules-based system in place to act as a check-and-balance to the predictive models. The beauty is as the predictive models identify common issues, those issues can be addressed with new rules, creating a stronger overall system than can be achieved with just one or the other.
No regrets
While the notion of “if I knew then...” may elicit nothing more than wistful regrets about personal choices, it can result in significant financeal losses to healthcare organizations if not identified and addressed quickly.
Predictive modeling technology can help healthcare organizations continuously discover what they don’t already know, or what they may have missed, so they can become progressively better at capturing all the revenue they’re due.
Author Bio: Paul Bradley, PhD, is the Chief Data Scientist at ZirMed, a recognized leader in cloud-based revenue cycle software and predictive analytics. He will be presenting on the topic of predictive analytics in healthcare at HFMA’s ANI conference on Tuesday, June 27th at 11am ET in ZirMed booth #1119.
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.