Picture of Rithvik Ramesh

Rithvik Ramesh

Socioeconomic Journal Club

Artificial Intelligence for Utilization Review |

“Of course it’ll depend on if your insurance okays it,” seems to resignedly asterisk many conversations in clinic. To attendings, I  am sure these considerations  are  normal  parts  of daily practice—some sort of balance between frustrating hurdles and valid checks on resource allocation. As a medical student having just recently started clerkships, though, they are some of the more glaring omissions from the neatly packaged paper cases of first-year. Utilization review, a process with which I am becoming increasingly familiar, ensures that patients receive necessary and appropriate care while also controlling costs for insurers and the healthcare system. However, with our nation’s staggering annual expenditure of over 1 trillion dollars on healthcare administration, we are in dire need of a more effective pipeline. Considering the growing potential for tools such as ChatGPT and automated scribes to improve clinical workflow, artificial intelligence (AI) is emerging as a key contender in the insurance space as well.

A recent JAMA Health Forum article by Mello and Rose at Stanford University explores the developing intricacies— legal, ethical, and practical—of health insurers’ budding use of AI for utilization review. Although the sheer volume of transactions and current inefficiency of claim assessment lend themselves to an AI-focused solution, the piece highlights class-action lawsuits and congressional inquiries complicating attempts at implementation. Allegations against UnitedHealthcare, Humana, and Cigna have embroiled these corporations in legal battles over their alleged misuse of blackbox algorithms. Specifically, complaints contend that AI using undisclosed criteria was exploited to determine ‘appropriate’ lengths of stay and restrict post-acute care exceeding these predictions. Not only were patients unable to access detailed information about these decisions, but more than 9 in 10 denials were reversed upon appeal.

The Centers for Medicare & Medicaid Services (CMS) responded to these AI experiments at the start of 2024 with a mandate. This rule allowed AI assistance for Medicare Advantage plan decisions on medical necessity with the stipulation that these determinations must account for “the circumstances of the specific individual” and should be “reviewed by a physician or other appropriate health care professional.” Of course, given sufficient room for interpretation and the relatively early stage of AI in utilization review, we can expect that medicolegal conflicts such as this are only just beginning.

The first major question in the wake of Mello and Rose’s article is what it means for these AI tools to incorporate patient-specific factors. Which variables best parallel human clinical intelligence and how dynamically should their weights adjust to accommodate the diverse values and experiences of our patients? As AI’s capabilities grow—transforming visits into notes, meticulously parsing charts, and generating a flood of metrics—the potential answers to this question will only multiply. Even without definitive solutions, Mello and Rose contend that advocating for transparency in algorithmic decision-making and the rationale for the factors behind coverage determinations is a crucial step in responsible AI usage. Especially as AI’s functions expand, we cannot allow the gap between algorithmic abilities and our understanding of them to become too large.

Another critical question concerns the accuracy of these AI insurers. Deciding what it means for these systems to be correct is inherently challenging, compounded by the reality that algorithms will never achieve perfection. Even an algorithm that is mostly correct will get thousands upon thousands of decisions wrong given our healthcare system’s volume. This inevitability raises the challenge of equity. When we look at existing health disparities and the growing body of work exploring structural biases in AI input and output, it is not hard to envision who will fall into the category of misses. If we train an algorithm on the current state of America, will we correct or will we amplify today’s differences in access and outcomes? Big data behind these algorithms is neither infallible nor complete. Moreover, the inherent financial interests in developing and tuning AI cannot guarantee that equity will be a focus of the process. As Mello and Rose aptly point out, even though it appears most denials can be appealed and reversed, we must also consider which groups disproportionately lack or possess the resources and know-how to navigate that process.

Still, having shared these reflections, it is impossible to reasonably suggest that AI should have no role in the future of this process. With the inefficiency of our current system and the enormous administrative burden on physicians, it would be irresponsible to not leverage AI’s capabilities. Furthermore, the use of predictive analytics to optimize decisions in our resource-limited reality is far from foreign to the insurance space, making AI a natural progression. Prior authorization, for example, represents a notoriously lengthy process that stands to benefit tremendously from the speed of advanced algorithms. So, the ultimate question really is not whether, but rather how AI should be integrated into utilization review.

As we will undoubtedly see reasonable guardrails put into place over the coming months and years, we should also consider the downsides of knee-jerk legislative responses to every attempt at AI implementation. Standardization and oversight of AI in utilization review require balance. Compliance with extensive regulatory frameworks can be painfully complex and expensive, with the potential to further magnify inefficiency. Furthermore, these regulations must contend with the uniquely rapid innovation in AI, which demands sufficient flexibility to adapt to emerging issues while also capitalizing on new opportunities.

Following the ethical and regulatory back-and-forth between physicians, patients, insurers, and the government will be important to everyone in healthcare—perhaps most to those of us still early in our careers at this important inflection point. As AI looms large, we should all think about what it means to have responsible algorithms and where in our practice, current or future, these technologies best fit. I, for one, welcome our new robot overlords—as long as they can help solve our   administrative issues.

Anthony DiGiorgio, DO, MHA

The key to AI regulation in healthcare is indeed balance, yet finding this equilibrium is no trivial feat. The concerns highlighted by the article regarding the opacity of algorithmic decision-making and the subsequent impact on patients are valid and demand attention. However, imposing stringent regulations could inadvertently stifle innovation and increase barriers to entry, disproportionately benefiting large institutions at the expense of smaller practices. A regulatory framework should foster transparency and accountability but must be flexible enough to encourage continued innovation and technological advancement.

Demanding that proprietary algorithms be made public as a solution raises significant concerns about the loss of competitive advantage and the hindrance of innovation. This open approach could discourage investment in new technologies, a scenario observable in other industries where proprietary technologies coexist with open-source models, allowing for diverse business strategies and innovations. Instead of mandating public disclosure of AI algorithms, a balanced approach could involve partial transparency. This method would protect proprietary information while ensuring sufficient oversight.

Furthermore, the impact of AI regulation on smaller practices and health equity cannot be overlooked. Over-regulation could lead to increased costs and complexities, sidelining smaller providers and potentially exacerbating health disparities among underserved populations. This must be considered when regulating AI, helping to ensure that AI technologies are both safe and beneficial before they are widely deployed.

In conclusion, crafting AI regulations requires a collaborative dialogue that includes a broad range of stakeholders from technology developers to healthcare providers, patients, and policymakers. By engaging all parties, we can develop a regulatory framework that not only protects patients but also fosters an environment where innovation thrives and is accessible to all healthcare providers, regardless of size. This inclusive approach will be crucial in leveraging AI to enhance health outcomes while maintaining a commitment to health equity and innovation.