The State of AI in Emergency Medicine

When I joined a Family Medicine practice in July 2004, fresh out of PA school, I was fortunate enough to share an office space with Dr. Robert Wettach, a 76-year-old GP who had spent decades caring for patients in a small town in southeast Iowa. Semi-retired, he still saw patients a few days a week, but his real passion was storytelling.

Between appointments, he regaled me with tales of medicine from his early days:  developing his own x-ray, mouth-pipetting hematocrits, making house calls, and being paid with baked goods or chickens.  He performed everything from appendectomies to tonsillectomies, delivered babies, and administered vaccines—witnessing the eradication of communicable diseases many of us have never encountered.  Notes were written in his own hand, and he later hired office staff to convert his dictations onto adhesive-backed paper stuck to a paper file.  He saw the advent of penicillin, the emergence of new specialties like emergency medicine, residency training, formal laboratory compliance requirements, the development of the CDC, and eventually computers and — inconceivably — the electronic medical record.  He didn’t care for it (or use it) much.

At that time, electronic medical records and voice-to-text dictation systems were relatively new, and not every clinic I trained in utilized one.  The vast majority of the small, rural clinics where I spent my clinical training years used some combination of paper charts, dictaphones, transcriptionists, and computerized records.  

Fast forward to my own career in emergency medicine. By 2009, I was navigating Epic for the first time. Over the years, technology has evolved rapidly: order sets, auto-populating templates, alerts for sepsis, transplant reminders, and inter-system EMR access. All tied to billing, coding, and traceable metrics. Now, 20 years after my first day as a PA, we’re on the cusp of yet another game-changer: artificial intelligence (AI).

The Future of AI in the Emergency Department

In a 2024 article in The Annals of Emergency Medicine, Dr. Robert Petrella, MD, suggests three stages that will allow AI to “give rise to profound changes in the field of emergency medicine.”  He outlines these stages, referring to their potential impact and the barriers and goals of each.  Ultimately, Patrella concludes that the introduction of and eventual everyday use of AI will affect many facets of ED operations, from patient flow to how we come to a diagnosis and how treatment decisions are made. There remains a paucity of data to validate these algorithms or support their widespread use before we fully adopt and implement AI’s use in these ED aforementioned operations.  

AI and Early Applications

The real-world application of AI in clinical practice is in its infancy.  We’re starting to see that patients prefer chatbot responses over physician responses to questions, rating them significantly higher for both quality and empathy.  AI algorithms exist for ED triage, some of which have been shown to assign ESI scores accurately.  Still, other studies suggest that current AI models are not an appropriate substitute for an expert triage nurse.  Further training and testing continues.   

Another novel use of AI in the ED that we may see become commonplace is AI assistance with X-ray interpretation.  Published in Chest Journal in July 2024, Rudolph et al. found an AI system outperformed non-radiology providers looking for pleural effusions, pneumothorax, consolidations, and nodules. Perhaps an AI model can more accurately identify a subtle pneumothorax on a chest x-ray than a non-radiologist. However, in a stable patient with normal vital signs, does that matter? 

AI’s potential is undeniable, especially as a tool to assist human judgment. Medicine is inherently humanistic, rooted in real-time decisions informed by patient history, physical exam findings, and intuition.

AI and Human Values

This leads us to another issue that arises when considering the application of AI to clinical medicine: AI’s lack of ability to consider human values and how those values influence a patient and provider’s risk-benefit consideration and reasoning.  

K.S. Yu et al. illustrate this concern in their paper entitled Medical Artificial Intelligence and Human Values, citing an example of an endocrinologist determining if a child with short stature should be prescribed recombinant human growth hormone.  They present clinical scenarios fraught with nuance regarding age, height, genetics, and poststimulation human growth hormone levels.  An AI model can be tasked with making treatment recommendations.  But whose values are guiding the algorithm? Clinicians? Insurance companies? Without careful oversight, biases in race, gender, or age could skew outcomes.

Anyone who has practiced medicine at the bedside has had conversations with patients regarding options for the workup of their chief complaint, the risks versus benefits of diagnostic imaging, and treatment options.  We review these decisions with our patients, recall evidence-based guidelines, and use shared decision-making while reading non-verbal cues and eliciting the patient's values, then select the most appropriate next steps.  These conversations “bridge the gap between objective data and personal values.”   Can AI models do the same?  We have not yet answered this question with a resounding “yes.”  

The authors conclude with an appeal to us all, “rather than replacing physicians, AI has made the consideration of values, as reflected by the guidance of a thoughtful physician, more essential than ever.”  Were Dr. Robert Wettach here today to see the advances of AI, particularly as it relates to its application to healthcare and patient care, I believe he would have agreed.