Beyond Data: Building a Clinically Intelligent Future
Joshua Tamayo-Sarver, MD, PhD, FACEP, FAMIA
I recall an instance where we tried having a large language model (LLM) summarize a patient’s anonymized chart, and it became fixated on a minor rash while overlooking a serious underlying inflammatory condition. Why? Because the rash, being the patient’s primary concern, received more documentation space and attention than the silent, but deadly, internal issue.
This experience underscored a critical need in healthcare AI: the need for a “clinical intelligence data layer.” We must move beyond simply feeding raw data into algorithms and ensure we develop new AI systems to understand the clinical context, the subtle cues, and the unspoken priorities that guide physician decision-making.
The Importance of Clinical Context
As healthcare professionals, we’ve spent decades trying to create a comprehensive data fabric that represents what’s happening in healthcare. Yet, we still struggle with interoperability and creating a data layer that can communicate across a patient’s care continuum. Even when we achieve this for a specific patient population, sharing that information across systems for population-level insights continues to be a challenge within healthcare AI.
In recent years, working on various AI products and tools, it’s become clear that we need an additional layer that contextualizes the information. This is particularly evident with LLMs attempting chart summarization. The problem with many new healthcare AI products being launched in this space is that they rely on how LLMs work — they recognize patterns and base their summaries on those patterns.
The Limitations of Pattern Recognition
Consider a patient with a serious inflammatory condition causing cardiac dysfunction, hypotension, and tachycardia. If the patient’s primary concern is a rash, that may be what gets the most attention in the documentation. An LLM, recognizing this pattern, might prioritize the rash in its summary, potentially neglecting a more serious underlying condition
This scenario highlights a crucial limitation in using LLMs: LLMs don’t inherently understand concepts of importance. They infer based on patterns, which can lead to critical oversights in healthcare contexts.
Building a Clinical Intelligence Data Layer
To address this challenge, we need to develop what I call a “clinical intelligence data layer.” This layer would mimic how a physician processes information — cutting through the noise, identifying important signals, and prioritizing them based on clinical significance.
Imagine an emergency physician rapidly assessing a patient, identifying immediate life threats, delayed life threats, and limb threats. This process of clinical prioritization is what we need to embed into our AI systems, which will require a hybrid approach of combining LLMs with other, non-pattern recognition, systems.
The Power of Contextually Appropriate Queries
With a clinically intelligent, prioritized data layer in place, LLMs could query this layer in a contextually appropriate way. This would enable them to produce meaningful summarizations and insights, assisting them in having a more accurate pattern recognition from which to infer whatever task is before them.
Looking Ahead: The Future of AI in Healthcare
As we continue to develop and refine AI technologies for healthcare, it’s crucial to remember that these tools are meant to augment, not replace, human expertise. The goal is to create AI systems that can understand and interpret healthcare data with the nuance and priority-setting abilities of experienced clinicians to support healthcare delivery in more efficient and effective ways.
By focusing on building this clinical intelligence data layer, we can unlock the true potential of AI in healthcare. We can create systems that not only process vast amounts of data but do so with an understanding of clinical context, leading to more accurate insights, better decision support, and ultimately, improved patient care.
The journey towards this clinically intelligent future is complex, but the potential benefits for healthcare providers and patients alike make it a worthy pursuit and exciting. As we continue to innovate in this space, let’s keep our focus on creating AI tools that truly understand and prioritize what matters most in patient care.
