A Focus on the Problem: A Balanced Approach to AI

At a time when artificial intelligence (AI) is often heralded as a panacea for all operational and clinical challenges in healthcare, a more balanced perspective is worthwhile. My recent experience at the HIMSS conference in Orlando exemplifies that while AI, particularly generative AI, holds transformative potential, its application within healthcare is still early and occasionally misguided. For example, in the case of prior authorization, one of my colleagues Dr. John Lee nicely stated, “are we just building a provider AI bot to argue with a payer AI bot?” Instead, shouldn’t we be streamlining the underlying rule-based approach to controlling resource utilization? A balanced approach to integrating AI into managing and optimizing electronic health record (EHR) system workflows must prioritize patient safety and clinician sanity, while ensuring we’re tackling the foundational challenges within our processes.

A Critical Eye on AI

The enthusiasm surrounding AI in healthcare is inescapable. I estimate 75% of discussions at the recent HIMSS conference focused on the integration of AI into various products and solutions. This widespread embrace comes with the caveat that not all AI solutions are created equal, nor are they universally applicable to every problem. The prevailing notion that AI is a catch-all solution can lead vendors and health systems to adopt technology that may not effectively address the underlying problems or, worse, introduce new complexities and liabilities. After all, the adoption of AI is just a tool in a toolkit, which comes with its own risks, benefits, and costs.

The Right Solution for the Right Problem

The distinction between a truly beneficial AI application and a marketing gimmick is crucial. My recent entrepreneurial crash course into product management has instilled the importance of problem identification before solution development. It’s essential to question whether the AI solution in consideration is genuinely a cost-effective and practical answer to the problem at hand. Will we see a Juicero in the healthcare AI space? In the space of EHR workflows, Aaron et al’s “Cranky comments: detecting clinical decision support malfunctions through free-text override reasons” publication found that a simple word matching heuristic was nearly as effective as more complex models in detecting clinical support system malfunctions; sometimes simpler is just as effective, while significantly easier to understand. This principle guides our approach at Phrase Health, where we remain skeptical of the “AI will solve it” mentality and instead focus on experimenting with AI where it makes sense to deliver clear, measurable value compared to other available approaches.

Experimenting with EHR Workflows

As a clinician dedicated to understanding EHR interventions and workflows, I (and the FDA) am acutely aware of the risks associated with deploying AI directly into patient and clinician workflows. Given our operational support role, we are lucky to work with the teams behind the scenes that support the front-line clinicians. As a result, our early AI development work has modeled validated approaches to known problems that exist for operational health system stakeholders. For example, we’ve done trials of deploying learning anomaly detection models to tackle the core problem of resource-constrained teams needing to simultaneously monitor 1000’s of alerts that are firing 24 hours per day, 7 days per week, and 365 days per year. More recently, we’ve actively explored using generative AI to benchmark order sets across our health system partners in order to identify outlier inclusion of inappropriate labs and medications. Our main consideration, as it should be with most AI solutions, is whether we can reduce the number of incorrect interpretations so as to not annoy our end-users.

Bringing It Together

Adopting AI in healthcare is a journey that requires a balanced, thoughtful approach. I recognize that I’m not alone as exemplified by the recent formation of the Trustworthy & Responsible AI Network (TRAIN). Several products showcased at the HIMSS conference highlight the industry’s inclination towards AI solutions, which are often interesting and exciting. Yet, it’s imperative to maintain a healthy skepticism towards claims of AI as a universal solution when the better approach may lie elsewhere. The path forward for Phrase Health involves leveraging AI where it truly makes sense and driven by a deep understanding of the core problems facing healthcare’s electronic workflows.