Disillusioned by AI in Medicine? There’s a Way Out, but It Might Be Messy

Young May Cha, MD

Machine learning and artificial intelligence (AI) emerged on the healthcare scene as a revolutionary technology that could leverage the massive data generated by medicine into sleek, personalized decision-making tools. But these technologies have yet to advance beyond the hype. Healthcare risks being stuck in the “trough of disillusionment,” a phase of the Gartner hype cycle, which describes the trajectory of innovative technologies.

Moderator Asad Siddiqui, BHSc, MD, MEd, FRCPC, assistant professor at the University of Toronto and a pediatric anesthesiologist at the Hospital for Sick Kids, opened this session “From Data to Decisions: How Data Science and AI Are Influencing Pediatric Perioperative Care,” which was co-sponsored by The Society for Technology in Anesthesia, on Friday, May 1 at the 2026 Annual Meeting, presented by IARS and SOCCA. The session delved into the challenges of developing predictive models and presented examples of how data science could impact pediatric perioperative care.

Despite the many published models, clinical decision-making remains relatively untouched by AI. To turn all the data into actionable insights, Theodora Wingert, MD, associate professor of clinical anesthesiology and clinical informaticist at the University of California, Los Angeles, discussed the need to tackle “the messy middle” that exists after the excitement of publishing a new model in which one must validate with real-time data and continue to do so after implementation. Unfortunately, perioperative pediatric care tends to have fewer and more heterogenous patients than perioperative adult care. Additionally, many aspects of pediatric medicine, such as developmental heterogeneity, age-dependent physiology, and measurement challenges, can make it more difficult to validate clinical prediction models. Even when models are high performing, they are often not clinically useful because the predictions come too late and the clinicians already know the patient is sick. A multidisciplinary team will be needed to help these models become truly useful for clinical practice.

Hannah Lonsdale, MBChB, associate professor in pediatric anesthesiology at Vanderbilt University Medical Center, continued this discussion by comparing the development of data science and AI in medicine to the development of new pharmaceuticals. Data science and AI need to progress past initial development into validation and calibration studies, impact studies, and implementation studies. For predictive models to succeed in these later steps, they need a good foundation that starts with a clinically important question and a large and relatively granular data set. Many models have high false positive predictive rates and will require extremely large datasets from multiple institutions to improve the positive predictive value. Although AI may be the trendy tool right now, sometimes conventional biostatistics may be the preferred tool. Dr. Lonsdale described using a logistic regression model to create the NEO-READY model to predict readiness for discharge in the NICU. Even after a model is created, further work is needed to implement it, validate it, and continue to validate it as temporal drift can occur within a model’s lifetime due to changes in clinical practice.

Tori Sutherland, MD, MPH, assistant professor of anesthesiology at the University of Pennsylvania, closed the session with an example of using national data to study teenage opioid use and build a screening risk algorithm to identify those at risk of opioid use disorder. Analysis of claims data in the United States showed that teenagers were prescribed more than twice the recommended opioid maximum for common surgeries, and the prescribing rates were significantly higher than other countries. Although these prescribing rates do not accurately reflect consumption rates, the diagnosis code “new persistent opioid use,” which is used for an opioid prescription 3-6 months after surgery, was used as a surrogate for chronic opioid use. There was a correlation between earlier preoperatively filled opioid prescriptions, older age (more than 15 years), preexisting chronic pain diagnoses (e.g., headache or abdominal pain), and mental health diagnoses (e.g., depression, anxiety, and substance use) with a higher rate of new persistent opioid use. Procedures found to be higher risk were lower pain procedures such as endoscopy, incision and drainage procedures, and laparoscopic procedures. These findings have led to a review of institutional opioid prescribing practices and greater education for trainees on opioid prescribing.

Machine learning and AI have not yet outperformed traditional methodologies, but they can be better for certain contexts, such as very large data sets or text data. But the work does not end with a successful clinical predictive model. These models will need to be analyzed at regular intervals to recalibrate them as needed because, once the model helps clinicians perform better, the new data can start making the model look like it is performing worse.