Interoperability and its technical and organizational challenges have remained largely unchanged since the early days of HealthIT. The emergence of Generative AI is revolutionizing the field, offering unprecedented opportunities to enhance data integration, streamline processes, and reduce costs.

This document explores how Generative AI is reshaping interoperability, impacting industry-standard APIs, Common Data Models, and ETL (Extract, Transform, Load) processes, and addressing long-standing complexities and inefficiencies.

The future of interoperability with AI is bright and exciting. In a series of posts, we examine the exciting new directions that Generative AI introduces, providing insights into its potential to modernize healthcare data exchange, improve accessibility, and drive operational efficiency. In this first post, we focus on taking advantage of Unstructured Data with more to follow.

The Real Cost and Challenges of Interoperability

Anyone who has been in the industry long enough can confidently agree that creating interfaces between systems is only give or take 10% technical effort and 90% bureaucracy and organizational logistics, such as compliance, approval, and other processes. If the one-time cost of putting initial integration infrastructure in place is taken out of the equation, this means that  the effort and cost of integrating with each new institution is largely the same - regardless of which major EHR they are running on and if we previously have integrated with any of these EHRs. That’s because bureaucracy and logistics are nearly entirely new and different for each organization. For a medical device manufacturer or software product vendor, which are looking to connect with many institutions, this is, of course, an enormous cost and undertaking. Moreover, since the purely technical side of the integration is relatively small, the variability in the interoperability of APIs and protocols between institutions and EHRs is also not a very significant cost and effort factor - provided these interfaces exist, of course.

Let’s explore how an AI-driven interoperability can help to address this unfavorable ratio.  

Taking Advantage of Unstructured Data

Healthcare practitioners have been dealing and working with unstructured content in the form of plain text, faxes, PDFs, and scanned documents since ancient times. Emergence and evolutions of EHRs and other technologies have been dramatically expediting the transition to structured data and management of it, but unstructured content in different forms (e,g, provider notes, discharge documentation, claims documentation) still play a critical role in the Healthcare workflows.

Using the latest GenAI models and related services and technologies, we have been developing solutions that are capable of extracting structured patient medical records from faxes, scanned documents, and other unstructured content with unprecedented precision and accuracy. The structured data may be extracted into a variety of formats, including Common Data Models such as FHIR and OMOP, as well as into proprietary representation. This data may be enriched and augmented with information from multiple sources beyond the original documentation and in other formats as part of this AI-driven extraction process.

Let’s take an example of a rehab facility that deals with patients discharged from hospitals. Most of these facilities in the US and around the world don’t have extensive IT infrastructure and resources to integrate with hospital infrastructures - integrating with many hospitals they receive patients from is out of the question for them. They receive most of the  discharge documentation as faxes, PDFs, and scanned paper and use a significant amount of manual labor to review, process, and transcribe these documents into whatever internal system they use (Point Click Care is a dominant player in the long-term care market). This work is not only expensive and laborious, but it may also lead to loss of business,  admission of wrong patients, and potentially critical errors.

A significant level of automation of such document acquisition and processing would mean the following for these facilities:

• No delays in making admission decisions - fewer lost patients and lost revenue to competition
• Improved quality and accuracy of patient records, leading to better quality of care and decision support
• Virtually no need for direct integration with many hospital systems, no dealing with organizational complexities and bureaucracies.
• The reliability, repeatability, and scalability of this workflow are far superior to the human processing. As one forward-looking multi-facility owner once told us, “AI doesn’t take days off and never comes to work in a bad mood”.
• Once in place, the automated processing of documents is a low-cost, ubiquitous IT task, far less expensive and complicated than manual processing and associated overhead. It can be easily outsourced if needed.

There are lots of similar use cases that we see nearly daily. They include processing, scrubbing, and validating claims for submission to payors, extracting insights from provider notes, prior authorization automation, validating FDA documentation against FDA guidelines, and a lot more.  

Of course, things are not as black and white as they may appear to be. Not all documents (like hand-written notes) can be currently fully and reliably processed by AI - or they require extensive additional work and training of the model. At least for now, some form of manual review is likely to be required. This manual work, however, requires far less resources and time compared to the existing fully manual processing. There are other factors and tradeoffs. These topics are addressable and should not be blockers for the overall automation of unstructured document processing.