C-Suite Q&A: ZirMed's Paul Bradley on Fixing Healthcare Payment

For the first time in our ongoing C-Suite Q&A series, Healthcare Analytics News speaks with a Chief Data Scientist. ZirMed is an analytics company trying to refine the process of provider reimbursement through the use of predictive modeling. Their chief data scientist is Paul Bradley, who spoke with us in July.

In Part 1 of a two-part interview, Bradley spoke at length of how the shifting healthcare industry has put higher payment burdens on patients, which has created the need for smarter systems to predict who will and will not pay bills. He quotes a colleague: “All of the goodwill that’s been accumulated when we have the patient in our care, that’s thrown out the window because we haven’t maintained the patient experience in the financial relationship we have with the patient.”

Rather than shuffling all unpaid balances to collections agencies, or getting stuck in payer limbo due to improper billing, health systems may be able to use good analytics to receive their due without aggravating patients.

Thanks for speaking with us. Can you first start with a short background of yourself?

My background is in databases, machine learning, statistics, data mining. I came through school with a PhD in Computer Science. After that I went to work for a division of Microsoft called Microsoft Research in Redmond, Washington. I worked with a group that, in addition to the research work we were doing around machine learning, we also transferred some technology to Microsoft’s flagship database product SQL Server. I’ve got a fair amount of experience with that system.

After I worked for Microsoft I consulted in a number of different areas which were not healthcare, a number of which were telco or ecommerce, and ended up in the revenue cycle space. We had hired a few people to do consulting work for us, and they said ‘you know, no one’s really doing predictive analytics in revenue cycle.’ So we thought we’d see what we could do, and that brings me to my position today.

What sort of players does your company work with in the healthcare sphere?

Primarily our clients are healthcare providers, hospitals and hospital systems, small doctor practices, long term care facilities, almost any place where healthcare is delivered. Our one sentence take on it is that “we help providers get paid.” We do a lot of work up front on pre-authorization and ensuring that insurance would cover a planned procedure from a patient. We then get data around the financial exchange between the payer and the provider after care is given, we do a lot of analytics around whether a claim has been properly coded to reflect the care that was provided to the patient. A lot of the predictive modelling work that my team does sits in that area, helping to catch coding errors or identify times when a code representing some billable item (a device or a drug or a procedure) is not on a claim. We’ve developed some machine learning technology to catch that.

We also have a claims processing clearinghouse, so we do process claims and we have that connection from providers to payers and we manage that reimbursement. Hence, we also have a lot of products around claim management and denial management, and those are also some areas where my team does some predictive modelling.

We tend to sit in the area of “revenue cycle management” but we’re kind of across the spectrum of prior to when care is given and after it’s given, making sure the coding is there, and even after reimbursement ensuring that it was made properly to the provider.

Is there any direct work with patients? What’s the impact there?

My team does a fair amount of analysis at the patient level, some of that is looking for care patterns around a given disease. Also on the financial end, more of the financial responsibility of healthcare is being put on the shoulders of the patients, through high deductible healthcare plans and other types of arrangements, so we’re hearing from more of our clients that they want some intelligence around this population of patients who owe them money. How do they interact with them?

I was giving a presentation at HIMSS 2 years ago with Richard Nagengast from Northwestern Memorial. Richard was talking about the amount of effort and focus that Northwestern Memorial puts into patient experience while the patient is in the care of the hospital: making sure that the patient knows what’s being done to them, what will be done to them, speaking to them in their native language, things like that. But when the patient is discharged, if it’s a patient like myself, a bill might come to me for the amount I know on Tuesday and it gets put in the stack of mail on the left side of my kitchen, and on Wednesday it goes to the right side, and it gets lost. Another bill comes and I don’t see it, and 4 weeks later I see a piece of mail come and it’s Northwestern turning me over to collections.

Richard made a good point in saying that ‘All of the goodwill that’s been accumulated when we have the patient in our care, that’s thrown out the window because we haven’t maintained the patient experience in the financial relationship we have with the patient.’ And then he went on to say that they’d leveraged some specific modelling technology to tell them the likelihood that the patient will pay their portion of the bill. If a patient is very likely to pay their bill, but just hasn’t paid yet, it’s way more beneficial for a provider to just wait for payment to come in for that patient, so they don’t have any ill will.

By turning it over to collections, you’re usually giving away a fair amount of that monetary return also. Patients who are unlikely to pay, there are opportunities to maybe find a government program that may cover their care, or see if they qualify for charity care. Patients in that middle spectrum, where you don’t have a lot of data either way to day if they will pay or not, maybe those are some you turn over to collections. It was a great example of ways to use analytics to help decide how to interact with patients who may owe a provider money for their care.

What is the reception like to that sort of thing among payers? Are there some that just prefer to defer to collections?

Not that I’m seeing with any of our clients, many of them are seeing more self-pay burden coming to them, with more patients burdened with paying either part or all of the cost of their care. In discussions with our clients, it’s often ‘how we do better communicate? How do we better interact with them in order to ensure payment or find a payment plan?’ I just feel like they’re looking for any information they can get to help them better deal with this population. The modeling that we’re doing is providing one pillar there to let them know what might happen with that group.

I wanted to go back to the analytics themselves, and the breadth of them. How much sheer data goes into any one specific case?

When I look at a given client of ours and how we model an encounter between a patient and a provider for one of our clients, we are typically extracting between 9,000 and 16,000 data elements around that encounter. Now, a given encounter doesn’t necessarily have all 16,000 elements populated, but that’s the type of breadth that we get. A lot of that is information about the patient, like gender, dates of service, how they were admitted to the hospital, also the diagnosis ICD-10 coding that summarizes that visit. What really helps us in the analytics is that we do have detailed charge data, which is at a lower level of granularity than you typically see on a claim. That allows us to find granular relationships between a certain type of, say, pacemaker and a certain type of drug.

We also do get the information on the attending physician, and we will see patterns. We do include the name or code for the attending physician as one of those data elements, and we do see patterns break out by attending physician. Typically this happens in orthopedics, where a certain doctor will use a certain set of bolts or screws to correct a joint, and another doctor will use a different set of devices. Having that low level of granularity allows us to, if I have a case where I know Doctor A was the attending physician, I have increased likelihoods of what devices he will use because I have analyzed historic data for Doctor A.

While the data is very granular and very wide, enough historic data allows us to find those detailed patterns and trends, and that allows us to really catch those 2% of times that providers aren’t properly coding for care that they’ve given.

Let’s talk a bit about data integrity. How do you go about ensuring that you can get the highest level of integrity in your data, and how do you validate the result from it?

One thing I do want to point at, at least when it comes to the charging data that is put onto a claim, typically across the industry hospitals are about 98% correct. 98% of the time, they’re not missing billable things being on the claim. As a data scientist, what that means to me is that there’s actually a fair amount of signal around the charging practices of a hospital or a provider’s location. What we do is we leverage the signal in that 98% of the data to identify these historic patterns and trends around charging behavior. To give an example: talking about charges for a pacemaker, I might have a cardiac event and I go into the hospital for a few days and I end up with a pacemaker. So 98% of the time, that pacemaker is correctly billed, and we can find patterns and trends around that, like what are the other drugs and procedures that have been given to similar patients in this scenario. We essentially start to build a profile of what common things occur when a pacemaker is provided to a patient.

There’s a lot of information in the data for us to derive that, and then we’re able to exploit it for the other 2% of the time. It typically that there’s usually a person in the process who is deriving codes to go in the claim, they’re either taking dictated notes from a doctor and looking to derive codes from them or they’re looking to text of a medical record. Because there is a person in this process, there’s just variance in how the coding happens. Analyzing this 98% signal in the data, we’re able to really zero in on cases where a number of the drugs and procedures that are often done in, say, that pacemaker procedure are on a claim, but actually that $15,000 charge for the pacemaker isn’t…the statistical models were raise a pretty big flag and say ‘this looks odd.’

Is there any background on the patient included?

We’ve been able to leverage some census-level aggregates, that are tied to the zip code level. We will know the zip code of the guarantor or of the patient who is responsible for care. In some cases, we have found that some of those demographic attributes summarized to the zip code level have provided us a bump. We really do attack this self-pay problem and all of our predictive modeling by analyzing all of the historic data a provider has collected. In the self-pay case, the biggest signal in that data comes from the historic payment patterns either from the patient or patients that are similar.

A number of providers address their self-pay population by getting a credit score for their patients, and based upon their credit score they will do certain things differently. We’ve had a number of examples where someone may have a very high credit score, but they will get a very low likelihood of paying their bill and when you look historically, they often haven’t paid their bills, and vice versa, someone with a low credit score might be a consistent payer of their healthcare bills. By analyzing historic payment patterns, we’re able to not be swayed in a different direction just based on a credit score.