Artificial Intelligence Documentation Prompts Must be Compliant, Too

Today I would like to share my opinion on proactive provider documentation decision-making technology. I am completely supportive of genuinely concurrent (that is, occurring in real time) clinical documentation integrity (CDI) efforts. However, I am afraid that the desire to leverage artificial intelligence (AI) can tend to push compliance to the wayside.

Think of the kind of dialogue that occurs when CDI specialists (CDISs) round with providers. Their goal is to generate verbal queries on the fly. When done compliantly, the CDIS shouldn’t lead the provider to alter their documentation; they should give them the necessary facts and clinical indicators, which permit the clinician to make a good, informed decision about best-practice documentation.

Why can’t we use technology to perform the same role? I think we can, but the programming that goes into the algorithm and offered choices needs to be done with care and in a compliant fashion.

I have spoken with representatives from multiple companies who are developing this type of technology, and they all resist the notion that these proactive, real-time documentation alerts are “queries.” The organizations that establish our CDI standards – the Association of Clinical Documentation Integrity Specialists (ACDIS) and the American Health Information Management Association (AHIMA) – are pretty clear that they do consider these queries.

In the ACDIS/AHIMA position paper on Guidelines for Achieving a Compliant Query Practice (2022 Update), they say, “the purpose and expectations of the documentation query process are to assist the provider in creating thorough and complete documentation, including specificity, treatment provided, and clinical validation. All queries must meet the same compliant standards, regardless of how or when they are generated, including those autogenerated by artificial intelligence (AI) and computer-assisted coding (CAC), whether in real-time computer-assisted physician documentation (CAPD) or after the episode of care is complete.”

There is also another ACDIS/AHIMA publication called Compliant Clinical Documentation Integrity Technology Standards, which asserts that any technology used to identify documentation opportunities must follow the guidance in the Guidelines for Achieving a Compliant Query Practice.

Some refer to these documentation alerts as “nudges.” This is defined by Merriam-Webster as “a slight push, poke, or jog (as with the elbow).” The word “leading” is defined in the dictionary as “guiding, directing.” Just by using the word “nudge,” it evokes the prohibited action of leading, because isn’t the point of a nudge to push the recipient in a predetermined direction?

Some real-time notifications are indisputably compliant. If a doctor has documented “heart failure,” instructing them to provide clarification as to specificity and type without providing specific choices is totally reasonable. The clinician has already established the diagnosis; the electronic CDI tool is merely asking for further detail.

If the computer is selecting clinical indicators and offering clickable potential diagnoses, it can get more complex. It needs to offer all diagnoses that can meet the applicable clinical indicators, and I would suggest that there should be a mechanism for the provider to reject all of them and/or explain their thought process. The ACDIS/AHIMA position paper on compliant querying states that “if a query response from a technology-driven query does not yield the response desired, it is inappropriate to send a follow-up manual query, for the same diagnosis/condition/procedure, in absence of new clinical indicators.”

If the provider just ignores the alert or isn’t given a choice of “I choose none of those offered,” how would the CDIS know that they shouldn’t manually query once they find the opportunity on their own review? The technology standards paper states that “all queries should be memorialized to demonstrate compliance with all query requirements and validate the necessity of the query.” Where are these documentation prompts with their instantaneous turnaround times memorialized?

I once saw a demonstration of one of these technologies, and the provider was offered the clinical indicators of an ejection fraction of 25 percent, with the patient having been administered a dose of diuretic, and the only choice offered was acute systolic heart failure. I pointed out that this could also be consistent with acute-on-chronic systolic heart failure, acute systolic and diastolic heart failure, and acute-on-chronic systolic and diastolic heart failure. Being given a single choice was misleading – or should I say, leading?

What if the technology noted an abnormal lab finding, such as a sodium of 131, and presented it as “an electrolyte disorder is noted. Please specify type.”? Again, this is not compliant. The provider has not already established that there is “an electrolyte disorder;” the e-CDIS is drawing a conclusion and making a diagnosis for them. It would be acceptable if the alert read, “there is an abnormal electrolyte level. Is there a corresponding diagnosis?” If asked my advice, I would actually recommend something like, “is there a clinically significant corresponding diagnosis?” and there would either be a reminder or provider training to document how the condition was being assessed, treated, or monitored. Without demonstrating clinical significance, it is not a codable diagnosis.

The Compliant CDI Technology Standards cites how computer-assisted provider documentation using AI and delivering prompts differs from information available for educational purposes on general CDI topics. The essential difference is that the prompt is case-based, and focused on details unique to each patient. This renders it a query.

There is guidance as to how to select a compliant CDI vendor in the Technology Standards paper. My advice is start with: “these are queries, and, as such, they need to be compliant.” Then, run with it. Everyone else is!