Medical language comprehension is a multi-headed beast. While computers have had the ability to scan medical records and extract key terms for many years, that alone is not sufficient. Multiple technical solutions must be implemented to effectively summarize clinical documents. Among them are OCR, dedicated language models for medical language and for scanned text, normalization, de-duplication, image processing, document segmentation, entity relation extraction, data enrichment, evidence linking, a robust and scalable backend, UI and more. This post will focus on a subtopic of data enrichment: concept relations.
There are many forms of concept relations. The most basic relationship is that between a child and a parent, where Concept A is related to Concept B if the proposition "A is a type of B" is true; we refer to this as an “is-a” relationship. Our teams at DigitalOwl have been manually curating those relationships for the last four years, covering almost two million relationships at present. The process of naively validating all potential relationships becomes impractical very quickly. In order to maintain such a large set of relations, a combination of medical expertise, statistical tools and computer code is required. These relations are imperative if one wishes to group the different manifestations of a disorder, such as spinal disorders for instance.
In Q2 of 2022, we began adding more complex relationships inspired by business workflows, underwriting manuals and industry experts. In order to analyze an impairment, an underwriter or claims analyst must be aware of not only the medical conditions for that impairment category ("is-a" relationship) but also relevant diagnostic or therapeutic procedures, lab results, medications, comorbid conditions, etc.
As this data is spread throughout the entire medical record, the underwriter or claims analyst had to jump from page to page to find all the relevant pieces of information. When utilizing the DigitalOwl solution, underwriters gain a 360-degree view of each key impairment so that the user has ALL the information they need to evaluate each impairment at their fingertips. This allows them to quickly evaluate the history without needing to review every page of the original APS.
As an example, given a history of cardiac disease, an effective summary will include all cardiac impairments together with additional details such as:
Cholesterol and eGFR from lab reports
Calcium Channel Blocker usage from the prescription history
Kidney disease from his most recent medical encounter
Angioplasty is retrieved from his hospital record.
Any specialist reports such as his cardiologist and nephrologist
All are neatly bundled together as part of the Cardiac Impairments section of our Digital Abstract.
It is our vision to have answers ready for every question an underwriter/claim adjuster might ask.
Here is a small sample of our data. The red dots represent conditions, the green dots represent medications and the blue dots represent procedures. The dots closer to the center are more general, and those on the periphery are more specific. Note the cross-domain relationships between conditions, relevant procedures and relevant medications.
As I mentioned, this is only a small sample. If you need a closer look at our data and how it will empower your team to make critical decisions faster, contact us today.