AMA brings forward a functional approach to classifying AI

Unless you are a medical coder, changes in the CPT code set are probably not that exciting to you. Many who work in healthcare don’t know that Current Procedural Terminology (CPT) codes are the universal language of the American health care system to describe the services and procedures provided to patients. In addition to being the basic language used in medical billing, these codes paint a picture of the type of care patients are receiving, from a basic sick visit in the office with their primary care provider, to the specific type of anesthesia provided during surgery, to the length of discussion and outcome of conversations about who is responsible for a patient’s medical decision if they are no longer able to make those decisions themselves. Like ICD-10, CPT codes are precise in their meanings, include guidance/prefatory language that is important in the application of the codes, and play a significant role in healthcare reimbursement.

In September 2021 the AMA released updates to Appendix S to describe work associated with AI-assisted medical services and procedures. The CPT Editorial Panel clarifies that the descriptions in Appendix S are applicable to all types of artificial and augmented intelligence (AI) applications used in clinical care, whether they are expert systems, machine learning, algorithm-based services, or other technologies. They also clarify that these classifications do not describe all work that is performed by machines in healthcare, but specifically they classify the work performed by the machine on behalf of the healthcare provider in the delivery of clinical care.

Appendix S describes three types of work that AI may perform on behalf of the healthcare provider:

Assistive AI: The work performed by the machine for the physician or other QHP (Qualified Healthcare Provider) is assistive when the machine detects clinically relevant data without analysis or generated conclusions. Requires physician or other QHP interpretation and report.

Augmentative AI: The work performed by the machine for the physician or other QHP is augmentative when the machine analyzes and/or quantifies data in a clinically meaningful way. Requires physician or other QHP interpretation and report.

Autonomous AI: The work performed by the machine for the physician or other QHP is autonomous when the machine automatically interprets data and independently generates clinically relevant conclusions without concurrent physician or other QHP involvement. Autonomous AI includes further classification based on the level of conclusions made by the AI and the role of the healthcare provider in response to those conclusions.

The AMA’s new AI taxonomy validates the Vitruvian Functional Framework for Workplace AI, which classifies AI by the impact it has on workers in performing and completing workplace tasks and decisions. In the Vitruvian Framework, AI can:

Accelerate: Increase worker efficiency in completing tasks and making decisions

Augment: Increase worker effectiveness in completing tasks and making decisions

Automate: Reduce worker effort in completing tasks and making decisions

Although the word choice is different between Assist and Accelerate, the meaning is the same because Assistive AI brings the relevant information to the physician or other QHP, which is designed to improve the clinician’s efficiency.

Importantly, the AMA brought forward a functional approach to classifying AI. They specifically emphasize that the type of algorithm used to generate intelligence is not important- it’s how the AI impacts a clinician’s tasks and decisions that matters most. The importance of this for AI adoption in healthcare cannot be overstated.

A functional approach facilitates designing solutions that appeal to end users, rather than designing them to conform to a mathematical or informatics construct. It enables AI designers, those tasked with selecting AI solutions, and those who are implementing AI technologies the insight needed to ensure the solution will appeal to the target end user. Clinical users of AI are not a monolith, and physicians, nurses, pharmacists, and social workers will not find the function of every AI solution equally appealing.

To learn more about how a functional approach to workplace AI can help drive appeal and create sustained adoption in healthcare or other industries, click here to download our paper.

Previous
Previous

How to Spot a Technology Unicorn

Next
Next

Can you address sdoh and disparities for $15?