It’s fair to say that the introduction of widespread artificial intelligence tools over the last four years has been one of the major changes of the time we live in. From its start, AI promised to change the world—we would be more efficient, more omniscient, able to answer any question with the click of a button.
And while AI has largely lived up to some of those promises, it has also introduced a set of risks and revolutionized a number of industries, not necessarily and certainly not entirely for the better. Unsurprisingly, one of these is Clinical Documentation Improvement software—in recent years, a number of companies, like SmarterDX and AKASA, have emerged on the market claiming to optimize pre-bill review and CDI through the use of AI, even though they are not integrated closely with the EHR workflows.
The field of AI is ever complicated because, like any new technology, it brings with it both advantages and disadvantages. In the CDI framework, it is often touted as a panacea. But when we take a closer look, is that really the case?
What AI does well in the medical sphere
AI has already started to shift the way we diagnose and research in the healthcare industry. Models are often able to find patterns and identify minuscule details on scans that the human eye simply can’t pick up.
This could even have the power to save lives. An April study from the Mayo Clinic used a researcher-developed AI model to detect pancreatic cancer from abdominal CT scans dated as much as three years before the actual clinical diagnosis. Looking at a sample of about 2,000 CT scans, the Radiomics-based Early Detection Model identified 73 percent of prediagnostic cancers at a median rate of about 16 months before diagnosis, per researchers.
Similar AI research at comparable medical institutions aims to flag cancer risk before patients may be diagnosed in a traditional clinical setting. These are tools that have the power to genuinely save patients’ lives and improve survival rates. It goes without saying that the earlier a patient is diagnosed with a life-threatening disease, the better.
How AI is affecting CDI
While early diagnosis is perhaps the most important way that AI affects the healthcare industry, it’s not the only way. The introduction of AI into the CDI process has already led to widespread changes, both on the provider side and the payer side.
In a recent interview with ICD10Monitor’s Talk Ten Tuesdays, Raemarie Jimenez, the president of membership and content for the American Academy of Professional Coders, AAPC, noted that AI wasn’t taking coder jobs so much as changing them. She referenced a recent study by Blue Health Intelligence, the data analytics arm of insurer, Blue Cross Blue Shield, that detailed how AI is sometimes misidentifying diagnoses.
In the study, released in March, data showed that from the second quarter of 2022 to the first quarter of 2025, acute posthemorrhagic anemia rates rose from 6.8 percent to 9.3 percent. This is a condition that generally indicates “severe blood loss and a need for extra medical attention and treatment, like a blood transfusion,” according to a press release from the Blue Cross Blue Shield Association. But transfusions present in admission, on the other hand, stayed nearly consistent across that entire time period. The increase in diagnosis of postpartum anemia constituted an additional $22 million in maternity admission costs in one year, per BCBSA.
“Something is disconnected,” said Dr. Razia Hashmi, BCBSA’s vice president of clinical affairs, in a press release. “Among hospitals showing the fastest rise in diagnoses of post-partum anemia, the rise in patients coded with this condition wasn’t paired with the level of care we would have expected, and the patterns we’re seeing point to AI-enabled coding.”
Healthcare providers that make the switch to AI-based tools have to understand that payers are also analyzing the data and looking for any gaps. As AI becomes more common, insurers will also become more astute at identifying issues in the data, like records that don’t thoroughly indicate pertinent negative results.
“Payers are analyzing our patterns and our coding data to find opportunities where care might not be reimbursed appropriately,” Jimenez said in the Talk Ten Tuesdays interview.
What are the ethical risks for CDI companies?
What makes AI coding tools so complicated is that they have the potential to work incredibly well—if supervised with clinician oversight. Medical coding is complex because it frequently changes and needs to be held to the most current research.
Our CEO, Dr. Gerasimos Petratos, was recently featured on the same podcast, Talk Ten Tuesdays, discussing the risks inherent with using AI in CDI software. His first point is perhaps the most important one to emphasize: an AI system is only as good as the data on which it is trained.
“An AI system trained on prior year codesets and not updated to reflect the new editions will inherently generate queries referencing outdated diagnostic terminology and miss newly available specificity options or fail to recognize that a previously singular code has been subdivided into multiple distinct subcategories,” he said in the interview.
This may seem simple—train the AI model on new information and it will stay relevant. But this is an issue built into the very system that claims to help us.
“These are not minor technical deficiencies, but structural inaccuracies in the clinical documentation clarification process that can result in miscoded claims and potentially false claims exposures,” Petratos said.
This is reinforced by the previous data referenced from BCBSA, as the same study also showed a rise in complex conditions claims from hospitals and healthcare providers that made a switch from standard coding systems to AI-based digital tools. In one case, there was an increase in roughly 6.7 percent of complex diagnoses after one provider made the change to AI. BCBSA estimates that AI tools can account for about $663 million in inpatient spending and $1.67 billion in outpatient spending. It also leaves the provider at risk for noncompliance claims and fraud disputes.
And just as AI needs to be updated to reflect new codes, it also needs to be modernized to reflect codes that are no longer in use. Retired and deleted codes that are not noticed as such by an AI CDI tool can result in providers coding for diagnoses that no longer exist and are now invalid.
How do we handle AI in CDI?
When it comes to new technology, the answers are never simple. While early adopters tend to dive in headfirst, thinking of themselves as innovators, that is not always the best strategy. Healthcare providers may feel like they are saving themselves a lot of money and getting better outcomes when they sign on to AI-based CDI systems that promise big savings and big returns on investment. But do they know what potential compliance issues might be tacked on to that?
The landscape around AI-based clinical documentation improvement is much more complex than it looks. Any company promising to earn a healthcare provider major Return on Investment (ROI) with contingency pricing probably requires a closer look. And as insurers become more aware of the possible violations in their midst, the regulatory ramifications will become increasingly more strict to navigate.
Our advice is to be on the right side of it now. At HITEKS, we’re using AI responsibly, concurrently and in keeping with compliance standards. If you’re interested in scheduling a demo and learning more about our product, click here.
