Can AI Unburden Physicians?

According to Arcadia.io , a cloud-based healthcare data analytics platform, “Hospitals produce an average of 50 petabytes of data each year with as much as 97% of that data going unused”. Moreover, they go on to say that “approximately 30% of the world’s data volume is being generated by the healthcare industry” [1].

Given this, it would seem that healthcare could benefit as much, if not more, from managing their data using AI than any other field!?

There is a clear need. According to the highly respected The Commonwealth Fund , our extraordinarily complex healthcare system is generally acknowledged to be far too error prone and inefficient. It is also, by far, the world’s most expensive system yet it produces some of the worse results as compared to its peer group of nations. [2]

One underlying reason is that already overworked physicians face ever increasing demands to document the quality of their patient care and provide information required by our uniquely complicated healthcare payment system. We turn now to what is probably the prime example of this.

The American College of Physicians (ACP) defines Prior Authorization (PA) as, “a common practice of health insurers in which physicians must first secure approval before moving forward with a patient’s medications, tests, or procedures in order to ensure the insurer covers that care”.? The college goes on to say that “Prior authorization is one of the most onerous administrative burdens that physicians face, forcing them to divert significant amounts of time and focus away from patient care”.

As the college explains it, PA “involves paperwork and phone calls, as well as varying data elements and submission mechanisms that can force physicians to enter unnecessary data in electronic health records (EHRs) or perform duplicative tasks outside of the clinical workflow”.?

Near the end of 2022 the Centers for Medicare & Medicaid Services (CMS), easily the single most influential organization in setting healthcare policies, issued a proposed rule titled in part “Advancing Interoperability and Improving Prior Authorization Processes”.? It calls for “Qualified Health Plan (QHP) issuers on the Federally-facilitated Exchanges (FFEs) to improve the electronic exchange of healthcare data and streamline processes related to prior authorization, while continuing CMS' drive toward interoperability in the healthcare market”.

The proposed rule calls for the implementation of a set of Application Programming Interfaces (APIs) it defines by the beginning of 2026.? More specifically it requires that “payers implement and maintain a Provider Access API that utilizes “HL7 FHIR version 4.0.1”. The API would facilitate the exchange of “current patient data from payers to providers, … adjudicated claims and encounter data … and the patient's prior authorization decisions”.

Might?generative AI help accomplish this?

In a?recent JAMA article?Dr. Robert-Wachter, author of the NY Times bestselling science book?The Digital Doctor?and?Professor and Chair of the Department?of Medicine at the University?of California, San Francisco, and a colleague said "we believe that genAI will deliver meaningful improvements in health care more rapidly than was the case with previous technologies."?Moreover, they go on to say “we see several reasons to believe that genAI will lead to productivity and/or quality gains more quickly than those achieved by previous tools and in previous eras”.? Despite the speculation as to whether or when AI will replace physicians in making diagnostic and treatment decisions, I’ve been arguing that its early and most impactful value to healthcare will come in improving and simplifying administrative processes -- like prior authorization.

I’ve previously written about GenHealth.AI and its unique approach to generative AI.? Unlike the Large Language Models (LLMs), such as ChatGPT, we’re all familiar with and which are trained on vast amounts of text information, the company’s Large Medical Model (LLM) is unique in that its trained on the structured, coded data charts and insurance claims of some 40 million patients. As a result, according to the company, their “LMM is more accurate in diagnosing medical conditions and recommending treatment options, while LLM is more proficient in understanding and generating human language”.

The company says it is working with health plans to automate and streamline the prior authorization process while ensuring compliance with upcoming CMS rule. They’re using their LMM to automate rule extraction from medical policies and create a simpler, faster approach to prior authorization.? The four-step process is illustrated here:

• ?A Data Crawler combs the insurance plan’s web pages, PDFs, faxes and other documents, and data warehouse abstracts, including those in the HL7 FHIR or Clinical Document Architecture (CDA or CCDA) formats.
• The GenHealth LMM transforms this medical policy into machine-readable rules and automatically builds forms and processes PA requests.
• A Rule Validation Administrative Apps allow staff to validate rules, produce PA reports, make manual interventions and modify AI developed logic.
• The company’s Prior Authorization API connects to the payor’s public FHIR server and executes the PA process as well as producing logs and reports.

Note that, as illustrated here, GenHealth requires the payer to provide a FHIR server so the entire process adheres to the HL7 DaVinci standards for FHIR-based Prior Authorization.

We will look at an example of the end result of this process – GenHealth’s AI derived Aetna Cardiac Devices and Procedures for Occlusion of the Left Atrial Appendage (LAA) Form for providers seeking PA approval on behalf of their patient.?

First, a bit of background. Atrial fibrillation (AF) is one of the most common cardiac arrhythmias. Its cause(s) are not yet clear, but it is associated with a number of other cardiac conditions, including in particular, diseases of the cardiac valves.? No matter its cause, individuals with AF have a higher stroke risk due to the increased possibility of thrombus (blood clot) formation. It is believed that thrombi could develop in the left atrial appendage (LAA) of the heart in certain individuals so it's suggested that closure by exclusion or occlusion of the LAA may reduce stroke risk in AF patients.

There are a variety of devices and clinical situations that impact on Aetna’s (or any other payer’s) willingness to pay for LAA occlusion. Aetna, a CVS Health Company , considers left atrial appendage closure (LAAC) devices medically necessary for non-valvular atrial fibrillation (NVAF) when the device has received FDA (U.S. Food and Drug Administration) Premarket Approval (PMA) for that device’s FDA-approved indication and the patient meets a list of some 18 criteria.

This is exactly the situation that the ACP is referring to when it says that PA ‘burdens the physician’.? What if, instead of having to figure this out, the physician could just answer a series of simple questions, ideally within their electronic health record (EHR) charting session using an EHR connected SMART on FHIR app?

That is the goal of GenHealth’s Prior Authorization API process, as illustrated here by the first five of the twelve questions derived using the company’s LMM to analyze Aetna’s left atrial appendage closure documents and turn them into rules.?

This is but one of a comprehensive set of AP questionnaires the company has posted for six payers, including CMS.? As illustrated, the web pages are interactive so you can choose a payer, the proposed treatment, answer the questions, and get a likely PA response from your chosen payer.

It is hard to imagine an application of AI with the potential for a more dramatic reduction in physician workload!

[1] https://arcadia.io/resources/taking-the-pulse-of-data-and-technology-in-modern-healthcare

[2] https://www.commonwealthfund.org/publications/issue-briefs/2023/jan/us-health-care-global-perspective-2022