It’s Time to Address the Potential Impact of AI on Diagnosis Selection

This week I thought I would revisit the concept of what provider documentation is. It is becoming increasingly difficult for clinical documentation integrity (CDI) and coding professionals to determine what’s actual, purposeful provider documentation, compared to what’s auto-populated into the health record. As ambient artificial intelligence (AI) scribes are adopted by healthcare entities and they become embedded into the electronic medical record (EMR) workflow, do we need industry guidance to redefine how we think about querying and the concept of leading?

Specifically, should technology companies be exploring the idea of managing patient diagnoses within the EMR? The foundation is already available due to regulations implemented during the adoption of electronic medical records within the Meaningful Use program. The goal was admirable: to require electronic health records to include a problem list as a way to promote collaboration through interoperability and improve patient care. The execution of this requirement, however, was problematic. Instead of being a tool that unifies patient care, the problem list often creates confusion regarding what diagnoses are related to a specific episode of care. Unfortunately, this ambiguity can extend into other areas of the health record when providers select predefined diagnoses within the EMR.

Rather than adopting ICD-10-CM when Stage 2 of Meaningful Use was implemented, Systematized Nomenclature of Medicine-Clinical Terminology (SNOMED CT) was selected as the common clinical language. In case you are not familiar with SNOMED CT, it provides core clinical terminology, with more than 300,000 concepts organized into hierarchies with formal-logic based definitions. According to the National Library of Medicine, it is the “most comprehensive multilingual clinical health terminology in the world.”

The dilemma is that when a provider selects a diagnosis within the EMR, they are really selecting a SNOMED CT concept that is mapped to an ICD-10-CM code. If you ask a provider, many will admit that it is overwhelming if presented with hundreds of possible diabetes codes, so they often may take the first one that appears (diabetes without complication), or maybe they will select diabetes with complication, but don’t realize they need to also specify the selection of another diagnosis. What happens is that we often end up with a problem list or even medical records with a variety of embedded diabetes codes that may conflict with each other. How is this helping anyone? The patient? The coder?

I must admit that I think it would be great if AI could maintain the problem list by automating a validation process, whereby if a provider begins to enter a diabetes code, they are asked if they want to use one already in the health record (i.e., confirm it), update the diagnosis (maybe the diabetic patient has developed associated complications), add an end date (e.g., AI could recognize diagnoses that are usually limited like an infection), or remove the diagnosis. This same approach could be used if providers must select a diagnosis within their assessment and plan – or elsewhere in the record.

Accurate diagnoses lead to better patient outcomes. The Agency for Healthcare Research and Quality (AHRQ) reported that “failures in diagnoses plague the general patient population across all settings of care.” The report continues by noting that “knowledge of the harms associated with missed, delayed, or inaccurate diagnoses is emerging.” Diagnosis failures cost healthcare $100 billion annually, according to AHRQ.

Accurately reflecting patient acuity is the role of CDI and coding professionals. We help providers translate their findings into ICD-10 terminology that best reflects their intent and represents the complexity of resources needed to treat each patient. However, hospital costs no longer include only those associated with patient care; administrative costs associated with the revenue cycle have skyrocketed, and can account for as much as 40 percent of hospital overhead. It is a bit of a vicious cycle, as payors add barriers, healthcare costs increase, and everyone loses.

But back to my original point about diagnoses in the EMR. There is inconsistency both within and across healthcare facilities in how CDI and coding professionals respond to embedded diagnoses. The issue is that the same technology that populates the problem list is also populating most of the diagnoses within the health record, because this is how the problem list is created in the EMR. This is also why some EMRs include the associated ICD-10-CM code with the embedded diagnosis.

Unfortunately, most providers do not have the time or skillset to validate the diagnosis that populates the EMR; most are fine if it is close enough. Fair enough, unless coders favor the embedded diagnosis code title over the associated narrative. For example, J18.9 pneumonia, unspecified organism, is often mapped to both community acquired pneumonia (CAP) and hospital (healthcare) acquired pneumonia (HAP). Neither identifies the type of organism that caused pneumonia, but documentation of HAP should prompt a query.

HAP is often associated with a more virulent pathogen, which will take longer to treat and require more resources; therefore, reporting J18.9 can result in revenue leakage, when assigned as the principal diagnosis. In fact, a 2020 article in the Cleveland Clinical Journal of Medicine (volume 87, number 10, p. 634) found that aspiration is an important contributor to HAP. The article also notes that there are limited diagnostic tests to verify this diagnosis beyond the findings of new lung infiltrates on chest imaging, respiratory decline, fever, and productive cough. This lack of a definitive test to identify the causative organism is also problematic because it requires providers to speculate on the cause of pneumonia based on how the patient responds to treatment.

The issue is that some CDI and coding professionals consider the embedded code title to be deliberate and accurate documentation of a diagnosis. Consequently, they will not or cannot (due to organizational policies) query the provider when there is clinical evidence of a more specific pneumonia diagnosis. One of the use cases for mapping SNOMED CT to ICD-10-CM codes, according to the National Library of Medicine, is “suggesting candidate ICD-10-CM codes to coding professional.” In other words, these embedded diagnosis codes should not preclude querying for more specificity when appropriate.

AI offers the healthcare industry a more efficient way to assist providers in determining the most accurate diagnoses, since it can synthesize clinical data within and across patient episodes of care. I am an advocate of exploring if AI can manage patient diagnoses by identifying when multiple similar codes are used to describe a diagnosis like diabetes with and without complications within the same episode of care.

The continued growth of ambient clinical documentation technology has the potential to proactively enhance provider documentation by translating it into ICD-10 coding terminology. The industry has already seen clinical “nudges” and other similar technology that prompts the provider to add more clarity if they document “CHF (congestive heart failure),” for example. What I can envision, if it is not already occurring, is leveraging AI to augment provider documentation by suggesting (not acting independently) modifications of selected diagnoses to fit the clinical scenario. If the provider documents CHF or if it is already on the problem list, AI could suggest chronic diastolic heart failure based on historically reported diagnoses, clinical findings, or both.

I am not trying to replace CDI with technology, but I do think we could do more to leverage AI in capturing the most accurate ICD-10-CM code when a provider is entering the diagnosis. I also think CDI would remain busy with reviewing the record for missing diagnoses, which is a much more complex process, and one that seems a little beyond the current capabilities of AI (or, at the very least, beyond the industry’s comfort level).

As AI becomes adopted into more EMRs, our industry needs to consider if and how we may need to modify our current practices, such as defining what a query or “leading” really is. Moreover, is the most important consideration the potential positive impact on patient safety by improving the accuracy of diagnoses within the health record?