To look more closely at industry efforts to streamline the collection and integration of complex medical data, RGA recently hosted a panel of key stakeholders to get their top considerations and perspectives on the formatting of underwriting data.
Data, the lifeblood of life insurance underwriting, arrives at an insurer’s doorstep in a tangled web of formats. Too often with structured underwriting data, form does not follow function. This troubling reality has limited insurers’ rates of automation, stymied their ability to glean critical insights from in-force policies, and obstructed the development of new underwriting methodologies.
As new vendors like DigitalOwl harness AI technologies to streamline complex medical data, the industry is also moving toward universal standards for structured underwriting data. To look closer at this exciting intersection, RGA recently hosted a panel of key stakeholders to get their top considerations and perspectives on the formatting of underwriting data.
Moderator:
Panelists:
Andy Kramer: When you think about structured data, think about data that can be easily analyzed – rows and columns of categorical data, like a metabolic panel or diagnosis codes. It's in a standard format. It can be analyzed in a decision engine or by actuaries or analysts. You can easily query any attributes you're looking for. On the flip side, unstructured data is difficult to analyze in an efficient manner, such as attending physician statements, the doctor's narrative about the patient’s symptoms. Our industry is trying to determine how can we turn unstructured data into structured data so we can use it in the underwriting process instead of having an underwriter go through that unstructured data and create a summary that can be fed into a decision engine.
One thing a lot of people don't realize is that when we get electronic health records, vendors typically take the raw data and apply their own style sheet to format it in a view that that an underwriter is used to seeing. But the real power is the underlying data. For example, when we look at attending physician statements today, they’re often print-offs from the doctor’s system. It's that same underlying data, just printed in the style sheet of their medical record software vendor. At the end of the day, when we talk about electronic health records (EHRs), it's essentially the same thing – just presented via a different style sheet. You may have a vendor and a style sheet that you like, but it's only scratching the surface on the capabilities we want to enable as insurance underwriters.
Dave Rengachary: A theme we're following when we talk about our needs around structured data is the question of “Who is using it, and what do they need it for?” Michael Hill manages optimization for RGA’s traditional facultative business. Michael, what really matters to you with structured data?
Michael Hill: First and foremost, it comes down to process efficiencies. If you look back at the past 10 to 20 years, we've seen a significant increase in the size of the cases referred to us – the number of pages of medical records has essentially doubled. That can be a good thing and a bad thing. On one hand, it's nice to have more data to allow us to more accurately assess risk. But underwriting review time has also increased.
We're also seeing an increase in the complexity of data. At RGA, we’ve partnered with AWS to leverage OCR (optical character recognition) and NLP (natural language processing) technologies to identify, extract, and structure data elements from scanned images. And we’ve partnered with DigitalOwl to pull in multiple data points to better assess the severity of medical conditions. DigitalOwl allows us to go beyond simple word recognition and begin to assess the severity of that condition.
We also have an opportunity to leverage structured data during audits, whether the audit takes place with the reinsurer or the carrier or is part of post-issue monitoring. With structured data, you can be more surgical in what you're looking for during audits. All roads around structured data lead back to operational efficiencies and time savings. - Michael Hill, RGA
Maria Beaulieu: I'm going to extend the conversation a little bit beyond just EHRs and APS (attending physician statements). When you're talking about underwriting and full automation, you're focused on taking all the various evidences together and looking at them in conjunction to assess a person’s medical conditions and individual risk.
When you look across the various evidences, you want the data to come in via a structure that allows you to compare and utilize all those evidences for that classification. You want the longitudinal view. If everything is in a different format, you’re forced to create rules and decision engines to assess that information. It requires extra effort to develop the technology and systems needed to process the data together to get to the end result.
Dave Rengachary: I'll turn to Nick next because he wears all sorts of hats for Northwestern Mutual. Whether it's from the perspective of a large carrier, or from your perspective of your work with reinsurers or other vendors, what matters to you when it comes to data?
Nick Milinovich: With any of these types of solutions, the table stakes are the human readable format. You can start with a style sheet and add on summary-type capabilities to streamline the underwriter’s focus. But even structured data is incredibly diverse. We see hundreds of thousands of different codes in EHRs. The dream is to distill those codes into impairments and disease entities, and then add in as much as you can from unstructured data. Then you can capture severity, symptomology, and other factors that are critical to the underwriting decision. It’s even better if we can put those findings into a clean package and use rules to adjudicate the majority of cases. Then our underwriters are free to apply their considerable skills to the most complex cases.
Dave Rengachary: We’ve talked about the use of structured and unstructured data during the acquisition of new business. But the same data can give us better insights on in-force policies. Mark will take us through that journey.
Mark Ma: In most companies, underwriting systems and actuarial systems such as experience studies are separated. After underwriting decisions, that rich underwriting information tends to be lost.
An underwriter may come to us, asking, “Hey, if I raise the cut-off point of an A1C result for preferred classes, how does that affect the business mix and mortality?” Embarrassingly, those questions are hard. They're hard because the wealth of underwriting data is locked in our underwriting systems.
As a result, we cannot quickly turn human experience into mortality insights. These limitations stifle my creativity as an actuary. - Mark Ma, RGA
Before an idea can mature, I automatically shoot it down: The idea is impossible since there is no data available. Digitizing underwriting files with DigitalOwl enables us to unlock a wealth of viable data that was previously inaccessible.
Each case is unique, yet with credible structured data and advanced analytics, we can move from case-level reviews to identifying the true drivers of the mortality experience. Actuaries can make better adjustments when migrating historical experience to a new targeted market or distribution channel. They can also change products with a high degree of accuracy and confidence.
Sounds great, but where do we start? Start with an impact study. Pull out a pool of applicants and digitize their underwriting evidence and applications. Take a look at the resulting digital summary. Does the summary match the underwriting decision? If not, is it because of additional impairments, the intricacy of comorbidities, or a lack of experience?
The inconsistency of human decision making can expose the complexity of considerations. The good news is that you have the data to build those considerations into your rules and models. By the end, you will understand the relationship between the features in a digital summary and your historical underwriting decisions.
Dave Rengachary: Let’s talk about the ongoing discussion about setting standards for structured data. Should you go with an existing healthcare-based format, or build an insurance-based format? Sean will talk about those different formats and the challenges you might encounter.
Sean Allen: The big question for our industry is, ‘How do you standardize things and put them in a structure so we can just look at mortality and get rid of all the morbidity issues?’ EHRs were built for clinical medicine; they’re intended for providers to pass information back and forth as they treat patients. In insurance, it’s a totally different perspective: We’re monitoring risk and trying to determine how people fit within our product sets. We want to do the kind of actuarial analysis and retrospective studies that Mark described. We’re also mindful of regulatory issues related to transparency and bias. How do we look at everything from one big picture?
Andy Kramer: Historically EHRs have been in proprietary formats because we as an industry never sat down to say, ‘Hey, let’s talk about standardizing this.’ At ACORD, we’ve spent nearly three years trying to drive standards to get more interoperability in EHRs. Hopefully everyone will adopt them so that we can achieve interoperability throughout the entire life insurance value chain, from production to reinsurance in the back end.
Nick Milinovich: Like any carrier, we at Northwestern Mutual have proprietary processes that apply to incoming data. We have proprietary manuals, automated rule sets, and predictive models. They're finely tuned to our market and our clientele, and they’re continually refreshed.
But many of the standards around data should not be proprietary, such as how the data comes in, the style sheets used by human underwriters, and the data structure that flows into a rule set. The problem around data is big enough that it needs to be solved as an industry.
Maria Beaulieu: Natively structured EHRs are dependent on the vendor applying their proprietary structures. For example, if you're getting a prescription record from Vendor A, it's not going to translate and read the same as a record from Vendor B. If we can’t all speak the same language, it’s difficult to complete risk assessment and put that data through automation. Carriers are spending their time building technologies to handle different iterations of the same information. That’s redundant and takes away from the time and effort we could spend addressing the information within the EHR.
Andy Kramer: One of the projects I'm working on is the standard application. One of our vendors said that if the industry could coalesce around one standard application, it would reduce the implementation cost of those electronic application systems by about 80%. And that’s because they’re spending 80% of their efforts mapping all these disparate application questions to the data structure on the back end. If we as an industry can get to standards, it will dramatically reduce the cost of adding new vendors.
Dave Rengachary: I’d like to switch gears and get an update from Andy about the mandates of ACORD and where that organization is in terms of moving the industry to standards.
Andy Kramer: ACORD is an international data standards organization for the global life and property and casualty insurance industry. We have two major EHR projects ongoing.
The first is around standards for EHRs. We started with the Fast Healthcare Interoperability Resources (FHIR) and Clinical Document Architecture (CDA) healthcare data standards. Then, we carved out everything not related to mortality, such as patient care and billing information.
I worked on one of the test files with Sean, in partnership with DigitalOwl and other entities. We started with a health information exchange record on a 71-year-old diabetic. The CDA record on this test individual was 25,000 rows of XML. After we applied the data standard, the file shrunk to 1,800 rows of HL7. That test case illuminates the reality that if our industry wants to get to point-of-sale approval – and possibly issue – of policies for complex cases, we can’t send 25,000 rows of code through our systems and expect a fast decision. But 1,800 rows can be analyzed very quickly.
ACORD moves the global insurance industry forward by encouraging and facilitating global information, enabling improved efficiency, and implementing effective, strategic positioning.
Both Maria and Sean have been instrumental in our second project, which is to take the top 200 impairments we see in underwriting and define the key attributes needed to assess an underwriting decision. With diabetes, for example, we’re drilling down and asking about Type 1 or Type 2, the date of onset, and measures like A1C. We’re also considering complications like kidney problems, vision problems, peripheral vascular disease, and coronary artery disease, and getting that into a structured data format. When applied, a vendor like DigitalOwl doesn’t need to send the 1,800 rows of data. They can just send the 20 fields that cover the key attributes for that impairment, which the carrier can then run through their RGA or other decision engine.
We’re nearing completion of the top impairments project; we’ve got about 115 of the top 200 completed. We’re making great progress.
Sean Allen: Where is DigitalOwl with standardization? We created our own standard data format, and I’ll tell you why.
People use data for different things. We started with the question of how to get APS data into a structured data format. We put it into an HL7 file format but realized very quickly its limitations, including around negated conditions. For example, was someone taking an antidepressant reactively because their spouse died, or was it part of a wider disease process?
This is where generative AI enters the picture. We basically created an API set called Connect. It can map to any decision engine. We use advanced AI and generative-text technologies developed exclusively for medical records.
By extracting medical information from the records submitted to the platform, DigitalOwl provides a focused summary of medical data points with an easy-to-use navigation system. The resulting reports ensure underwriters can access the most meaningful data points in hours versus days. Underwriters can use this for new business underwriting, in-force analysis, or any other underwriting methodologies they’d like to explore.
- Sean Allen, DigitalOwl
Dave Rengachary: I’ll end by mentioning that the standard RGA has landed on is the medical standard developed by the DigitalOwl team. Their standards allow more complex cases to run through automated underwriting systems. They also provide the next level of evidence optimization and strengthen the invaluable feedback loop around in-force assessments. The solution will also be “ACORD-aligned” to complement this emerging and important industry standard. Our team will lead by example; we have a great front-row seat on DigitalOwl’s transformation of data structure.
