The following article was originally published in On The Risk.
Effective underwriting has always hinged on the ability of underwriters to interpret and organize vast amounts of data, from medical records to diagnostic findings. This data structuring in underwriting has traditionally been a labor-intensive, manual process requiring underwriters to sift through hundreds of pages to gather the relevant information necessary to make sound decisions.
However, with the advent of AI-assisted tools, the potential for streamlining the way underwriting data is structured and utilized is evolving. Generative AI models can consolidate and organize complex information, making it more accessible and actionable.
But this begs the questions: How do these advancements compare to traditional methods… and what do they mean for the future of underwriting?
Drawing on RGA’s partnership with DigitalOwl, this article explores real-world use cases to examine the opportunities and challenges of leveraging AI to structure and present underwriting data.
Let’s examine an underwriting case: A 71-year-old woman presents with a history of anxiety, depression, type 2 diabetes, and chronic pain requiring opioid medications. In reviewing this file, we determined that while her diabetes was well controlled, her depression had worsened and was associated with concurrent anxiety, and she had recently started on a new medication, as shown in the images below.
Using this information, the underwriter accurately assessed this applicant’s risk. However, this underwriting file totaled approximately 433 pages, including the application, paramedical exam and labs, and medical records outlining the applicant’s medical history.
Thanks to RGA’s collaboration with DigitalOwl, underwriters now have access to their generative AI models that streamline the process by creating PDF summaries. These summaries transform lengthy underwriting files, such as this 433-page case, into a more concise, accessible format as shown in Exhibit 3.
Duplicate information that was initially found scattered across multiple areas of the underwriting file is now presented in a concise, underwriter-friendly format. This allows the underwriter to navigate the applicant’s medical findings strategically and logically, enabling more efficient and confident risk assessments.
Generative AI’s ability to structure medical and diagnostic findings makes the data in each underwriting file easily digestible across various contexts and formats.
For example, in addition to PDF summaries, DigitalOwl produces both Excel output files and machine-readable JSON files at the case level. Their DATA API technology further enhances deeper analytics by allowing users to retrieve large volumes of structured data through API calls based on medical items, mentions and more.
For clients requiring customized outputs for batch processing, there are two key solutions available:
The Excel output file below provides the same information initially summarized in the PDF format. However, it is now organized by relevant impairments, dates each impairment occurred, and even potential severity as predetermined by the AI’s configuration.
With the ability to formulaically classify various “medical items,” underwriters can now prioritize medical findings by relevance and severity. In addition to the medical items listed above, the Excel output file also contains the following information:
This overview represents only the surface of what is possible. When diving deeper into the data within these files, new pathways emerge, leading to previously untapped opportunities.
For example, the same depression history outlined in the medical records and PDF summary can also be found in the machine-readable JSON format, shown in Exhibit 5. Through generative AI’s ability to categorize each medical finding, all potential medical codes associated with these findings are also provided, such as LOINC, ICD-10, SNOMED, etc.
Organizing these impairments and findings in this manner makes it feasible to input data into various rules engines or severity models to streamline underwriting decisions. In other words, the same information that previously required an underwriter to manually review a 400-plus page file can now be interpreted by a rules engine.
Clearly, the impairment analysis has become easier and more efficient with the integration of AI. But, again, we have only begun to explore the possibilities of consolidated and organized data.
Take laboratory data, for example. Given the variety of sources in which lab results are provided – whether from insurance panels, third-party data, electronic health records, or attending physician’s statements, underwriters must often review each independent lab value from its respective source. This process can be cumbersome, requiring underwriters to piece together results to estimate mortality based on recent findings, trends, or averages.
However, with generative AI’s ability to re-structure all lab findings, we can now develop rules- based logic that is more encompassing. The following snapshots illustrate how lab tests can be categorized, filtered, and assessed, regardless of their source.
Whether using a customized tabular worksheet or a case-specific Excel output, underwriters can now consolidate laboratory findings from all available sources into a streamlined format. This consolidation offers significant opportunities:
Furthermore, categorizing these findings by “key impairments,” and associating them with relevant codes, such as the A1c test highlighted in green in Exhibit 6, enhances data consolidation. This enables the development of impairment-specific rules or severity models. With all underwriting data consolidated, underwriters gain the ability to organize relevant information logically, dynamically, and accessibly, unlocking a wealth of possibilities.
In essence, the value of structured data lies in its organization. By presenting information in an intuitive, easy-to-consume format, structured data empowers underwriters to achieve more precise, efficient, and informed decision-making.
Like PDF summaries, structured data is just a way to present the results of a generative AI model. As such, concerns about accuracy in AI models also apply to the structured data they produce.
Validating this data is critical. For example, the model may misinterpret or omit medical findings, especially when the input data is complex or unclear. Exhibits 8 and 9 illustrate a case where the model incorrectly identified an individual as having diabetes mellitus. Further investigation revealed that this was simply a diagnosis code used to order glucose testing, and the test result itself was normal. Nevertheless, this erroneous diabetes diagnosis appeared across all structured data formats as well.
Exhibit 10 highlights another challenge: A patient’s Crohn’s disease was omitted because the generative AI model was unable to reliably interpret the handwritten details. As a result, all structured data formats failed to mention the patient’s history of Crohn’s disease.
The presentation of medical impairments or diagnostic findings in structured data comes with unique challenges. For instance, while Exhibit 6 highlights the significant benefits of the customized tabular worksheet for data consolidation, there are many instances where fields like “Institution” are left blank. Underwriters must evaluate the relevance and reliability of assessments based on the information source. Further refinement is needed for generative-AI to accurately identify and associate medical findings with their exact sources.
These challenges are due to ongoing efforts to validate structured data based on PDF summaries and an underwriter’s ability to review them. DigitalOwl’s PDF summaries provide hyperlinks to source data, enabling underwriters to verify the legitimacy of the data. This allows underwriters to trace impairments back to their original sources and assess what led to the condition’s “diagnosis,” as seen in the diabetes example in Exhibits 8 and 9.
Consistency and accuracy in structured data are essential – not only for proper risk management in underwriting but also for auditing and claims adjudication.
As generative AI continues to evolve, whether applied to underwriting workflows or case notes, like those from DigitalOwl, the need for improved validation tools and techniques will become critical to advancing structured data analysis and risk assessment.
With continued validation and refinement, the challenges of ensuring accuracy and building confidence in AI-structured data will gradually diminish. Once these hurdles are overcome, opportunities for structured data consolidation and rules development will expand, turning what once seemed a distant future inevitability into a present-day reality filled with possibilities.
And this is only the beginning.
Contact us to learn more about how RGA and DigitalOwl are pioneering AI-enhanced underwriting workflows that drive efficiency and better risk decisions.
