The Adoption of Artificial Intelligence in Clinical Care

As a newly minted division chief at a major children’s hospital at the beginning of the 2000s, I (FWP) observed the growing impact of the electronic medical record (EMR) movement on clinical care. Now, as a wizened graybeard, I’m wondering about the effects of artificial intelligence (AI).

Lessons From the Electronic Medical Record Movement

In my opinion, much of the promise of EMRs has yet to be realized. The widely hyped potential of EMRs to simplify the transfer of medical records from one institution to another is often thwarted by the fact that, even when nominally the same EMR model is implemented by the same vendor at two separate institutions, one hospital’s version is incompatible with the other’s. This is a result of the “customization” of the EMR to accommodate each institution’s special needs and policies.

Furthermore, as I consult on cases, I note that although the patient’s symptoms have changed profoundly over the course of a hospitalization, the clinical notes fail to reflect these changes accurately. Instead, outdated excerpts from previous clinical notes are often repeatedly “cut-and-pasted” into the progress notes, sometimes perpetuating serious errors.

Finally, when we attempt to transmute an EMR’s bits and bytes into a paper record, it often proves unusable. For example, every blood pressure, pulse, lab value, or medication administration is printed on a separate page, resulting in a haystack of paper that defies organization.

The Clinical Promise of Artificial Intelligence

AI is inexorably coming to clinical care and will be far more impactful than the EMR movement to date. AI is often promoted as improving diagnosis and prediction of the clinical course. It is increasingly being used to inform decision-making and patient management. AI is also used to review patient records and generate notes and discharge summaries. Finally, AI is emerging as an important tool in research and teaching.

A number of studies find that AI can equal or exceed the diagnostic accuracy of experienced specialists—in a fraction of the time. AI’s capacity to process enormous amounts of data is unmatched. For example, a recent study seeking to predict the likelihood of agitation and violence in emergency rooms examined more than 3 million visits. Interestingly, the findings were “…consistent with existing literature that identified historical violence as a predictor of future agitation events” (Wong et al., 2025). Thus, in this case, AI confirmed the longstanding clinical truism that past behavior is the best predictor of future behavior.

AI can also “see” behaviors that are difficult for clinicians to recognize. For instance, identifying a seriously ill premature infant’s state of arousal is extremely important for the optimal timing of feeding and care (Putnam, 2016). Cell phone videos of infants in a neonatal intensive care unit were analyzed with an AI object detection model quantifying head and hand movements (Nishio et al., 2025). The results provided an easy way to identify the best times to care for desperately ill preemies.

Problems With Artificial Intelligence

With the rapidly growing use of AI, a number of serious problems have emerged that affect a broad range of applications. So-called “hallucinations” are false “factual” statements made by AI programs. These may include invented references and even fabricated data. Disturbingly, the frequency of hallucinations seems to be increasing with each new version released (a peculiarity common to all AI platforms to date) and may occur as frequently as 40 to 75 percent of the time.

Old-timers, such as myself, worry about an uncritical overreliance on AI by clinicians as well as misrepresentations of risks and benefits to patients. Indeed, some studies suggest that AI-based health care searches may generate massive amounts of disinformation, including convincing, but AI-fabricated, medical images (Menz, 2023).

What Is Being Done to Improve the Clinical Uses of AI?

Much of the pioneering development of EMRs occurred in the Veterans Administration (VA) with its 170 hospitals, about a decade before the rest of the medical community began adopting EMRs. Now, drawing on its extensive experience with EMRs, the VA has become a leader in using AI to develop clinical predictive models as well as integrating their results into routine patient care.

One lesson is that AI predictive models drift over time and probably should be recalibrated on an annual basis. Another is that AI models are poor at predicting low-frequency medical or mental health events (e.g., suicide), leading to disruptive false alarms that vastly outnumber true positives.

Although many AI models focus on clinical prediction, one of their more important functions is simply searching a patient’s EMR for relevant symptoms and linking them to the appropriate scientific literature. The VA is pioneering technologies such as “ambient listening,” which uses AI to record, transcribe, and analyze patient-clinician conversations, automating documentation and freeing the clinician to focus on the patient and not the keyboard.

Stephan Fihn, one of the pioneers in the VA’s work on EMRs and an early adopter of AI, believes that the future of AI processing of a patient’s EMR will look more like a Google search rather than a traditional medical chart—pulling together all of the relevant information with a single click (Perlis, 2024).

If so, AI may help the EMR to finally realize its true potential.